一句话：提出以假设为中心的LLM根因分析框架HERO，通过信念跟踪与歧义引导证据采集，显著提升微服务故障诊断准确率。
云原生微服务系统的根因分析需要整合异构可观测信号，并在有限诊断预算下决定下一步检查。现有基于大语言模型的RCA智能体要么依赖预定义工作流，其效果受限于不完整的标准操作过程覆盖，要么采用开放式诊断，在多个合理根因假设与观测证据一致时表现脆弱。本文提出HERO，将诊断形式化为竞争假设间的受控判别，通过有界假设空间实例化、显式信念状态跟踪和歧义引导的证据采集来指导诊断。作者在OpenRCA和真实工业故障数据集上与代表性工作流式及开放式基线对比。在OpenRCA上使用DeepSeek-V3时，HERO达到79.77的成功率，超过最强ReAct式基线23.20个百分点，且平均少用1.04个动作；在工业数据集上达到85.00成功率并保持最佳。结果表明，显式控制竞争假设是微服务RCA中有效且实用的抽象。
Root-cause analysis (RCA) in cloud-native microservice systems is challenging because it requires integrating heterogeneous observability signals and deciding what to inspect next under limited diagnosis budgets. Existing large language model (LLM)-based RCA agents either rely heavily on predefined workflows, whose effectiveness is constrained by incomplete Standard Operating Procedure (SOP) coverage, or adopt open-ended diagnosis, which remains brittle when multiple plausible root-cause hypotheses are consistent with the observed evidence. We present Hero, a hypothesis-centered RCA framework that formulates diagnosis as controlled discrimination among competing hypotheses via bounded hypothesis-space instantiation, explicit belief-state tracking, and ambiguity-guided evidence acquisition. We evaluate Hero on OpenRCA and a real-world industrial incident dataset against representative workflow-based and open-ended baselines. On OpenRCA with DeepSeek-V3, Hero achieves an SR of 79.77, outperforming the strongest ReAct-style baseline by 23.20 points while using 1.04 fewer actions on average. On the industrial dataset, it reaches an SR of 85.00 and remains the best-performing method. These results suggest that explicitly controlling competing hypotheses is an effective and practically promising abstraction for microservice RCA, especially in real-world, production-style incidents with incomplete workflow coverage and high diagnostic ambiguity.