一句话：提出约束推理链语法 CRCG，将专家告警推理形式化为可检查的显式推理链，仅在验证通过时保守削减告警，实现在真实数据中心场景下零误减。
告警洪水是系统运维的关键难题：单一故障会在互连组件间引发级联告警，淹没真正的故障信号。告警削减以安全为导向，任何错误削减都可能漏掉真实故障。现有方法存在根本性权衡：规则系统可靠但难以扩展，基于大语言模型（LLM）的方案泛化强但具有非确定性、推理不可验证，难以用于安全关键场景。本文研究数据中心运维中的告警洪水削减问题，提出基于结构化推理与保守验证的方案。其核心洞见是专家级告警分析依赖连接上游原因与下游告警的显式推理链。作者设计了约束推理链语法（CRCG），一种轻量且具表达力的语法，将此类推理过程形式化为带逻辑与数值约束的组合推理链。方法分两步：先用 LLM 辅助标注、检查器引导编译与监督微调（SFT）将专家知识组织为 CRCG 并训练模型生成显式推理链；随后仅在推理链可依 CRCG 验证通过时才保守削减告警。在真实数据中心 HVAC 告警数据上的评估表明，该方法在评测设定下取得高准确率并将误减降至零，优于规则与纯 LLM 基线，并显著降低运维工作量。
Alarm floods pose a critical challenge in system operations, where a single fault can trigger cascaded alarms across interconnected components, overwhelming human operators and obscuring actionable signals. Alarm flood reduction aims to reduce redundant alarms under a safety-oriented objective, because any incorrect reduction may falsely exclude real failures. Existing approaches face a fundamental trade-off: rule-based systems ensure reliability but lack scalability, while solutions based on a large language model (LLM) offer strong generalization but suffer from non-determinism and unverifiable reasoning, limiting their use in safety-critical settings. In this paper, we study the alarm flood reduction problem in data center operations and introduce a solution based on structured reasoning and conservative verification. Our key insight is that expert-level alarm analysis relies on explicit reasoning chains that connect upstream causes to downstream alarms. We design Constrained Reasoning Chain Grammar (CRCG), a lightweight yet expressive grammar that formalizes such reasoning processes as compositional inference chains with logical and numerical constraints. Our method follows two tightly connected steps. First, we use LLM-assisted annotation, checker-guided compilation, and supervised fine-tuning (SFT) to organize expert knowledge in CRCG and train a model to generate explicit reasoning chains. Second, we conservatively reduce alarms only when the reasoning chain can be verified based on CRCG. We evaluate the method on real-world data center HVAC alarms. Results have shown that our method achieves high accuracy while driving false reductions to zero on the evaluated setting, outperforming both rule-based and LLM-only baselines and reducing operational workload substantially.
@inproceedings{paperbot12334,
title = {Towards Reliable Alarm Flood Reduction in Data Center via Grammar-Guided Verifiable LLM Reasoning},
author = {Jiacheng Liu and Guocong Lin and Jichao Yan and Huiwen Zheng and Hongli Ma and Changhao Zhou and Zhiyi Yu and Yuanchun Li and Yunxin Liu and Ding Li and Tao Xie and Yao Guo},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834481}
}