摘要：提出了一种基于大语言模型（LLM）进行蒙特卡洛树搜索的智算网络故障根因分析系统RCA-MCTS。利用LLM推理研究领域在蒙特卡洛树搜索上的前沿研究，面向智算网络复杂故障场景，设计了适用于故障根因分析任务的多策略提示语扩展机制，并基于与故障模拟环境交互反馈的方式设计了模拟机制，使得LLM推理时的蒙特卡洛树搜索过程适配于故障根因分析任务场景。实验表明，RCA-MCTS在故障根因分析任务准确率上提升33%～43%，在故障推理动作序列平均匹配度上提升18%～34%。

关键词：智算网络；故障根因分析；大语言模型；蒙特卡洛树搜索

Abstract: A fault root cause analysis system of intelligent computing networks based on Monte Carlo tree search (MCTS) and large language models (LLM), named RCA-MCTS, is proposed in this paper. By leveraging cutting-edge research on MCTS in the domain of LLM reasoning, a multi-strategy prompt expansion mechanism is designed for fault root cause analysis tasks in intelligent network fault scenarios. Additionally, a simulation mechanism is developed based on feedback interactions with the fault environment, enabling the MCTS process during LLM reasoning to be adapted to the fault root cause analysis task. Experimental results show that RCA-MCTS improves the accuracy of fault root cause analysis by 33%-43%, and enhances the average matching degree of fault inference action sequences by 18%-34%.

Keywords: intelligent computing network; root cause analysis; large language model; Monte Carlo Tree Search