Can LLMs only talk? Experimental studies on task scheduling with Large Language Models
Mengjuan Li, Zhengguang Chen, Huan Zhou, Zhipeng Wang, Yingwen Chen, Baokang Zhao, Xue Ouyang, Jinshu Su · 2025
Large Language Models (LLMs) have emerged as a disruptive technology for Natural Language Processing (NLP), achieving success in NLP-related generative applications. However, the potential capability of LLMs in other domains remains largely unexplored. To explore the potential of task scheduling with LLMs, we model a typical task scheduling scenario in cloud computing and transfer scheduling problems as natural language prompts. Afterward, the knowledge and reasoning abilities of LLMs are enabled to generate scheduling decisions. Six well-known and open-source LLMs are integrated into our framework to perform experimental studies, and the results are evaluated from multiple perspectives and compared with each other. Besides, traditional heuristic algorithms and a basic Reinforcement Learning (RL) method are all performed for comparison. Our results demonstrate: 1) compared to most heuristic methods, the decisions made by LLMs achieve better scheduling performance; 2) compared to the basic RL method, LLMs exhibit better generalization on various workload patterns; 3) the larger parameter size of the LLMs has, the better scheduling performance it achieves. To the best of our knowledge, our experimental study is the first exploration to apply LLMs in task scheduling. Our findings highlight the promising potential of LLMs as a novel approach to task scheduling, offering new avenues for research and practice.