一句话：提出AIM框架，用LLM统一多模态告警数据，生成摘要、根因分析与可执行缓解脚本，并评估云端与本地模型。
面向大规模微服务云系统中告警过载、上下文碎片化和故障恢复延迟的问题，现有方法很少给出可执行恢复步骤或语义摘要。本文提出 Agentic Incident Mitigation（AIM），一个自适应、LLM 驱动的框架，统一多模态日志、指标、追踪与告警数据，生成简洁告警摘要、根因分析洞见和可执行缓解计划；并将缓解计划作为提示，交由代码生成 LLM 通过上下文学习与自适应提示选择生成修复脚本，无需昂贵微调。实证比较云端 GPT-4o 与本地模型，自适应提示使 RCA 准确率和缓解质量在 ROUGE 与余弦相似度上提升 7–10%，同时降低 token 成本。LLM 生成 playbook 的任务完成率为 71.9%，但二值修复成功率为 25%，可靠性受数据集特性影响。
Site Reliability Engineers (SREs) face growing challenges in maintaining resilient cloud systems, as large-scale microservice deployments generate continuous streams of logs, metrics, traces, and alerts. Alerts are critical for Root Cause Analysis (RCA), but their volume often causes fatigue, fragmented context, and delayed fault resolution. Existing approaches rarely produce actionable recovery steps or semantic summaries. We propose Agentic Incident Mitigation (AIM), an adaptive, LLM-driven framework that unifies multimodal data to generate concise alert summaries, RCA insights, and executable mitigation plans. AIM further uses mitigation plans as prompts for code-generation LLMs to produce remediation scripts, employing in-context learning with adaptive prompt selection to improve reasoning and output quality without costly fine-tuning. In an empirical study comparing cloud-hosted (GPT-4o) and on-premise models, adaptive prompting improved RCA accuracy and mitigation quality by 7–10% across ROUGE and cosine similarity metrics while reducing token cost. Using Task Completion Rate, TCR, we found that LLM-generated playbooks achieved 71.9% task completion, though binary remediation success was 25%, with dataset characteristics affecting reliability. AIM transforms raw alerts into interpretable RCA narratives and executable actions, establishing a human-in-the-loop, AI-guided, privacy-preserving paradigm for intelligent cloud fault management.
@inproceedings{SardaJZWSL26,
title = {Leveraging LLMs for Alert Summarization and Mitigation Plan Generation},
author = {Komal Sarda and Honggeun Ji and Amr M. Zaki and Ian Watts and Larisa Shwartz and Marin Litoiu},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3805255}
}