RuntimeSlicer: Towards Generalizable Unified Runtime State Representation for Failure Management

RuntimeSlicer: Towards Generalizable Unified Runtime State Representation for Failure Management Lingzhe Zhang Peking University Beijing, China zhang.lingzhe@stu.pku.edu.cn Tong Jia ∗ Peking University Beijing, China jia.tong@pku.edu.cn Weijie Hong Peking University Beijing, China hongwj@stu.pku.edu.cn Mingyu Wang Peking University Beijing, China mingyuwang25@stu.pku.edu.cn Chiming Duan Peking University Beijing, China duanchiming@stu.pku.edu.cn Minghua He Peking University Beijing, China hemh2120@stu.pku.edu.cn Rongqian Wang Huawei Technologies Co., Ltd. Beijing, China wangrongqian2@huawei.com Xi Peng Huawei Technologies Co., Ltd. Hong Kong SAR, China pancy.pengxi@huawei.com Meiling Wang Huawei Technologies Co., Ltd. Shenzhen, China wangmeiling17@huawei.com Gong Zhang Huawei Technologies Co., Ltd. Shenzhen, China nicholas.zhang@huawei.com Renhai Chen Huawei Technologies Co., Ltd. Beijing, China chenrenhai@huawei.com Ying Li ∗ Peking University Beijing, China li.ying@pku.edu.cn ABSTRACT Modern software systems operate at unprecedented scale and com- plexity, where effective failure management is critical yet increas- ingly challenging. Metrics, traces, and logs provide complementary views of system runtime behavior, but existing failure management approaches typically rely on task-oriented pipelines that tightly couple modality-specific preprocessing, representation learning, and downstream models, resulting in limited generalization across tasks and systems. To fill this gap, we propose RuntimeSlicer, a unified runtime state representation model towards generalizable failure management. RuntimeSlicer pre-trains a task-agnostic rep- resentation model that directly encodes metrics, traces, and logs into a single, aligned system-state embedding capturing the holis- tic runtime condition of the system. To train RuntimeSlicer, we introduce Unified Runtime Contrastive Learning, which integrates heterogeneous training data sources and optimizes complementary objectives for cross-modality alignment and temporal consistency. Building upon the learned system-state embeddings, we further pro- pose State-Aware Task-Oriented Tuning, which performs unsuper- vised partitioning of runtime states and enables state-conditioned adaptation for downstream tasks. This design allows lightweight task-oriented models to be trained on top of the unified embed- ding without redesigning modality-specific encoders or preprocess- ing pipelines. Preliminary experiments on the AIOps 2022 dataset demonstrate the feasibility and effectiveness of RuntimeSlicer for system state modeling and failure management tasks. CCS CONCEPTS • Software and its engineering→ Maintaining software. KEYWORDS Failure Management, State Representation, Metric, Trace, Log 1 INTRODUCTION Modern software systems continue to evolve at an unprecedented scale and level of complexity. From e-commerce platforms and so- cial media services to financial trading systems and cloud-native infrastructures, today’s software systems routinely support billions of users and execute business-critical workloads under highly dy- namic conditions. However, the widespread adoption of microservice architectures, containerized deployments, and elastic scaling mechanisms has significantly increased the dynamism, heterogeneity, and unpre- dictability of runtime environments. As a result, large-scale soft- ware systems experience failures more frequently, leading to sub- stantial operational disruptions and financial losses [5,30]. Ac- cording to an ITIC report, unplanned IT downtime in production environments costs large enterprises over one million U.S. dollars per hour on average [12]. These challenges highlight the critical need for timely failure management, including anomaly detection, failure localization and failure classification. System traces, metrics, and logs capture rich information about the runtime behavior of software systems, recording both the inter- nal states of executing components and the critical events occurring during system operation. As such, they constitute data sources for failure management, providing insights into normal system behav- ior as well as deviations that may indicate emerging failures. A large body of prior work has explored failure management techniques based on these runtime data modalities. Metric-based approaches analyze key indicators of system performance and resource utilization—such as latency, throughput, and CPU or memory usage—to detect anomalies or infer potential failure arXiv:2603.21495v1 [cs.SE] 23 Mar 2026 Conference’17, July 2017, Washington, DC, USALingzhe Zhang et al. Trace Metric Log Preprocessing Metric Encoder Trace Encoder Log Encoder interpolation downsample... 3-sigma log parsing tokenization merging ... critical path extraction DAG construction ... Task-Oriented Model Anomaly Detection Failure Localization Failure Classification ..... Trace Metric Log Runtime Slicer Task-Oriented Model System-State Embedding Model Training Model Training Figure 1: Comparison between conventional failure man- agement training pipelines and the RuntimeSlicer-enabled unified training workflow. causes [3,16,17,19–21,25,38]. Log-based methods leverage system logs as sequences of structured or semi-structured events, com- bining temporal patterns and semantic information to uncover abnormal behaviors and failure signatures [4,8–10,14,22,29,31– 35]. Trace-based approaches focus on service invocation relation- ships and request propagation paths across distributed compo- nents, and are commonly applied to failure localization and failure classification tasks in microservice systems [2,6,15,27,28,42]. More recently, multimodal approaches attempt to jointly leverage logs, metrics, and traces to improve failure detection and diagno- sis by fusing complementary runtime signals from different data sources [11, 13, 18, 23, 24, 26, 36, 37, 39–41]. The effectiveness of existing failure management methods has been demonstrated. However, as illustrated in Figure 1, current ap- proaches still suffer from fundamental limitations. Specifically, most existing methods adopt a task-oriented pipeline: metrics, traces, and logs are first processed through modality-specific preprocess- ing procedures, followed by separate encoders tailored to each data modality. The resulting representations are then fused and fed into task-specific models, such as anomaly detection, failure localization, or failure classification models. During training, the modality encoders and task-oriented models are optimized jointly in an end-to-end manner. This tightly coupled design causes failure- related patterns to be implicitly entangled with task objectives and encoder architectures, making the learned representations highly dependent on specific tasks, datasets, and system configurations. As a result, the extracted features are difficult to reuse across tasks or environments, leading to limited generalization when system behaviors, workloads, or failure patterns change. To fill this gap, we propose RuntimeSlicer, aunified runtimestaterepresentationmodeltowardsgeneralizablefailure management. RuntimeSlicer pre-trains atask-agnosticrepresen- tation model that directly ingests metrics, traces, and logs, and encodes them into a single, aligned system-state embedding cap- turing the holistic runtime condition of the system. Unlike conventional task-oriented pipelines that jointly train modality-specific encoders and downstream models, RuntimeSlicer decouples representation learning from failure management tasks. Once trained, RuntimeSlicer can be applied to different systems to produce system-state embeddings, on top of which lightweight task-oriented models—such as anomaly detection, failure localiza- tion, and failure classification—can betrainedwithoutredesigning preprocessingpipelinesorre-trainingmodalityencoders. To train RuntimeSlicer, we introduce Unified Runtime Con- trastive Learning, a representation learning framework that inte- grates diverse training data sources, including labeled datasets, run- time collection from live systems, and controlled failure injection. The training objective combines multiple complementary loss func- tions, including modal consistency loss for cross-modality align- ment, temporal consistency loss for preserving runtime continuity, and an optional weak anomaly loss to incorporate coarse-grained failure signals. Building upon the learned system-state embeddings, we fur- ther observe that system states are naturally structured and recur- rent, corresponding to distinct runtime conditions such as work- load levels, traffic patterns. Motivated by this observation, we pro- pose State-Aware Task-Oriented Tuning, which first performs unsupervised state partitioning over system-state embeddings to identify latent runtime states. For each identified state, RuntimeS- licer employs a state-aware adaptation mechanism to train state- conditioned task models, enabling downstream failure management tasks to adapt to heterogeneous runtime conditions. We conduct preliminary experiments on the AIOps 2022 dataset [1], which is collected from a mature microservices-based e-commerce system, to evaluate the effectiveness of RuntimeSlicer in distinguishing system runtime states and supporting failure man- agement tasks. 2 METHODOLOGY RuntimeSlicer-enabled failure management is organized into two main stages: system-state embedding generation and state-aware task-oriented tuning. As shown in Figure 2, the left part of the pipeline illustrates the embedding generation process, where met- rics, traces, and logs collected within the same time window are jointly fed into a pre-trained RuntimeSlicer model. RuntimeSlicer, trained via Unified Runtime Contrastive Learning, encodes these runtime signals into a compact system-state embedding. Once the system-state embedding is obtained, it can be used as input to the State-Aware Task-Oriented Tuning stage, which sup- ports downstream failure management models. It is worth noting that the primary goal of this paper is to pre-train a generalizable runtime state representation through RuntimeSlicer. The proposed state-aware task-oriented tuning serves as one possible instantia- tion demonstrating how the learned embedding can be exploited, while the representation itself is compatible with a wide range of downstream tasks and model designs. 2.1 Unified Runtime Contrastive Learning Unified Runtime Contrastive Learning is designed to learn a unified and task-agnostic representation of system runtime states from heterogeneous observability data. Unlike task-oriented training RuntimeSlicer: Towards Generalizable Unified Runtime State Representation for Failure ManagementConference’17, July 2017, Washington, DC, USA Time Time Window Runtime Slicer Traces Logs Metrics Offline Unified Runtime Contrastive Learning System-State Embedding RuntimeSlicer Embedding GenerationState-Aware Task-Oriented Tuning Anomaly Detection Unsupervised State Partitioning State #1 State #2 State #3 State #n ...... State-Aware Adapter State-Conditioned Task Models 1 2 3 n ... Failure Localization Failure Classification Density Model Deviation Score Normal Abnormal Component Encoding Contribution Score Root Cause Rank Failure Prototypes Relative Encoding Failure Type Figure 2: Pipeline of RuntimeSlicer for failure management. The left side illustrates how RuntimeSlicer, trained via Unified Runtime Contrastive Learning, ingests metrics, traces, and logs to produce a system-state embedding. The right side shows how this embedding is leveraged for State-Aware Task-Oriented Tuning of downstream failure management models. Traces Logs Metrics T Pretrained Embedding Model Fusion Layer Representation Model Modal Consistency Loss Temporal Consistency Loss Weak Anomaly Loss W1 W2 W3 Abnormal1 Normal1 Normal2 Normaln Abnormal2 Abnormaln (Optional) Loss Objectives Training Data Labeled Dataset Runtime Collection Controlled Injection Figure 3: Training Pipeline of Unified Runtime Contrastive Learning paradigms that entangle representation learning with specific fail- ure management objectives, our goal is to pre-train a representation model that captures the intrinsic structure of runtime behaviors and can be reused across different systems and downstream tasks. As shown in Figure 3, Unified Runtime Contrastive Learning consists of three key components: training data construction, representation model, and loss objectives. Training Data Construction. To support robust and scalable representation learning, we leverage three complementary sources of training data: (1)LabeledDatasets.We utilize publicly available datasets, such as AIOps 2022 [1] and ART [24], which provide synchronized met- rics, traces, and logs along with failure annotations. These datasets offer explicit supervision and serve as a reference for learning failure-aware runtime representations. (2)RuntimeCollection.We deploy a set of open-source dis- tributed software systems, including Train-Ticket [43] and Online Boutique [7], under diverse workload configurations. During exe- cution, we continuously collect metrics, traces, and logs generated by the running systems without requiring manual labeling. This data source enables RuntimeSlicer to capture the intrinsic structure and natural variability of system runtime behaviors. (3)ControlledInjection.Building upon the runtime collection setup, we further introduce a limited amount of fault data through controlled failure injection using Chaos Mesh. This data source is intentionally kept small, as learning a unified runtime state repre- sentation does not fundamentally rely on explicit anomaly labels. Instead, injected failures are used to provide weak supervisory sig- nals that enhance the model’s ability to distinguish coarse-grained abnormal and normal states. Conference’17, July 2017, Washington, DC, USALingzhe Zhang et al. Representation Model. RuntimeSlicer is instantiated by inte- grating a shared pre-trained embedding backbone with minimal ar- chitectural adaptation for unified multimodal representation. Given a time window푡, we denote a multimodal runtime observation as Equation 1. 푥 푡 =(푥 푀 푡 ,푥 푇 푡 ,푥 퐿 푡 )(1) In the equation,푥 푀 푡 ,푥 푇 푡 , and푥 퐿 푡 correspond to metrics, traces, and logs respectively. RuntimeSlicer adopts a unified pre-trained embedding backbone푓 휃 (Qwen3-Embedding-0.6B in our implemen- tation), shared across modalities, to encode each input into a latent semantic embedding, as illustrated in Equation 2. 푒 푀 푡 = 푓 휃 (푥 푀 푡 ), 푒 푇 푡 = 푓 휃 (푥 푇 푡 ), 푒 퐿 푡 = 푓 휃 (푥 퐿 푡 ).(2) Unlike conventional architectures that design modality-specific encoders independently, RuntimeSlicer leverages a shared linguis- tic–statistical embedding prior to ensure that heterogeneous signals are mapped into a unified embedding space, enabling natural com- parability and modality-aligned reasoning. The resulting modality embeddings are aggregated via a light- weight fusion layer푔 휙 , implemented as a shallow multi-layer per- ceptron (MLP) with optional gating as Equation 3. 푧 푡 =푔 휙 Concat(푒 푀 푡 ,푒 푇 푡 ,푒 퐿 푡 ) (3) The fusion layer finally producs a compact system-state embed- ding푧 푡 ∈ R 푑 that summarizes the holistic runtime condition of the system, reflecting workload fluctuations, temporal dependencies, service interactions, and emergent failures. Loss Objectives. Unified Runtime Contrastive Learning jointly enforces three complementary constraints to shape the system-state embedding space. (1)ModalConsistencyLoss.To ensure that metrics, traces, and logs reflect a coherent system condition within the same time win- dow, we align modality-specific embeddings using an InfoNCE-style contrastive term. For a batch of time windows푡 푖 퐵 푖=1 , let푧 푀 푖 ,푧 푇 푖 ,푧 퐿 푖 denote metric-, trace-, and log-level embeddings respectively. The loss aligns each modality pair by treating same-window represen- tations as positives and others as negatives as Equation 4, where sim(·)denotes cosine similarity and휏is a temperature parameter. Losses for(푇,퐿) and(푀,퐿) pairs are computed analogously. L modal =− 1 퐵 퐵 ∑︁ 푖=1 log exp sim(푧 푇 푖 ,푧 푀 푖 )/휏 Í 퐵 푗=1 exp sim(푧 푇 푖 ,푧 푀 푗 )/휏 (4) (2)TemporalConsistencyLoss.Runtime states exhibit smooth temporal evolution, where adjacent time windows sharing sim- ilar runtime spans are expected to remain semantically close in the embedding space. Let푠 푖 be the state embedding of window푡 푖 , and휔 푖푗 ∈ [0,1]denote their temporal-overlap ratio, which softly weights the expected similarity as Equation 5, where훿is a slack margin and푍is a normalization constant. This formulation natu- rally induces smoothness for temporally coherent runtime periods while avoiding excessively forcing embeddings of unrelated win- dows to collapse, thereby preserving meaningful separation when temporal evidence suggests divergence. L temp = 1 푍 ∑︁ 푖≠푗 휔 푖푗 max 훿− sim(푠 푖 ,푠 푗 ), 0 (5) (3)WeakAnomalySeparationLoss(Optional).When coarse anomaly labels exist, we introduce a light constraint to prevent abnormal states from becoming overly similar to normal ones. Let NandAdenote normal and abnormal sets of embeddings as Equa- tion 6, where훾controls the maximum permitted similarity. Notably, this term does not cluster anomalies, which only avoids accidental entanglement between normal and abnormal states. L anom = E 푠 푛 ∈N,푠 푎 ∈A max sim(푠 푛 ,푠 푎 )−훾, 0 (6) 2.2 State-Aware Task-Oriented Tuning Since RuntimeSlicer fundamentally captures the underlyingsystem state, we further enable downstream failure management to become state-adaptive. LetS=푠 1 , . . .,푠 푁 denote the learned system-state embeddings. We first perform unsupervised state partitioning as shown in Equation 7, identifying latent runtime regimes such as workload tiers, traffic conditions, and resource-pressure states: C=퐶 1 , . . .,퐶 퐾 , 퐶 푘 =푠 푖 ∈S | assign(푠 푖 )= 푘.(7) To operationalize such structure, we introduceState-Aware Adapter, a lightweight module that conditions downstream task models on cluster-specific contextual information. For a downstream failure-management taskT, we derive a set of State-ConditionedTaskModels 휃 1 , . . .,휃 퐾 , each tuned on sam- ples belonging to one state cluster as Equation 8 which yields state- specialized models while maintaining a shared global backbone. 휃 푘 = arg min 휃 E 푠 푖 ∈퐶 푘 L T (푥 푖 ;휃,푠 푖 ) (8) For different downstream tasks, RuntimeSlicer instantiates the above mechanism as follows:AnomalyDetection—state- conditioned density models estimate deviations from typical behav- ior to produce anomaly scores;FailureLocalization—state-aware component encodings yield contribution scores for root-cause rank- ing; andFailureClassification—state-specific prototypes enable rel- ative encoding–based classification within each state regime. 3 PRELIMINARY EVALUATION To evaluate RuntimeSlicer, we conduct a preliminary study on the AIOps 2022 dataset to assess its feasibility. We first examine the quality of RuntimeSlicer’s learned represen- tations by evaluating its ability to distinguish runtime system states. As illustrated in Figure 4, we compare 2D t-SNE visualizations of runtime-state embeddings generated by Qwen3-Embedding-0.6B versus RuntimeSlicer. The baseline embeddings exhibit no clear structure, with points densely mixed across the space, whereas Run- timeSlicer yields well-separated clusters, indicating that it captures latent system-state patterns more effectively. This suggests that RuntimeSlicer provides state-aware representations that are more suitable for downstream operational tasks. RuntimeSlicer: Towards Generalizable Unified Runtime State Representation for Failure ManagementConference’17, July 2017, Washington, DC, USA 402002040 Component 1 30 20 10 0 10 20 30 40 Component 2 Normal Abnormal (a) Qwen3-Embedding-0.6B 302010010203040 Component 1 30 20 10 0 10 20 30 Component 2 Normal Abnormal (b) RuntimeSlicer Figure 4: Comparison of runtime-state embeddings gener- ated by Qwen3-Embedding-0.6B and RuntimeSlicer Next, we evaluate its performance on downstream failure man- agement tasks. We report Precision, Recall, and F1 for all tasks; ad- ditionally, for Failure Localization and Failure Diagnosis, we also re- port Mean Reciprocal Rank (MRR). As shown in Table 1, RuntimeS- licer combined with State-Aware Task-Oriented Tuning achieves promising performance across all three failure-management tasks. However, we observe performance degradation in cases where certain runtime states are underrepresented in the training data. We plan to focus on optimizing performance for such scenarios in future work. Table 1: Failure Management Results TaskPrecisionRecallF1-ScoreMRR Anomaly Detection97.27%81.18%88.50%- Failure Localization69.57%67.33%68.43%70.15% Failure Diagnosis87.57%75.12%80.88%83.35% 4 CONCLUSION This paper presents a unified runtime state representation for gener- alizable failure management, eliminating the need for task-specific encoder training. To this end, we introduce Unified Runtime Con- trastive Learning, which jointly leverages heterogeneous labeled and unlabeled data to learn system-state representations, yield- ing RuntimeSlicer. Based on these embeddings, we further pro- pose a demonstration framework—State-Aware Task-Oriented Tun- ing—to showcase how RuntimeSlicer supports downstream failure- management tasks. Preliminary experiments verify its feasibility, and future work will focus on improving generalization through large-scale fine-tuning. ACKNOWLEDGMENT This work was supported by the Huawei–Peking University Joint Laboratory of Mathematics. REFERENCES [1] 2022. AIOPS 2022 Championship. https://competition.aiops.cn/. [2]Yang Cai, Biao Han, Jinshu Su, and Xiaoyan Wang. 2021. Tracemodel: An auto- matic anomaly detection and root cause localization framework for microser- vice systems. In202117thInternationalConferenceonMobility,Sensingand Networking(MSN). IEEE, 512–519. [3]Jiaxiang Dong, Haixu Wu, Haoran Zhang, Li Zhang, Jianmin Wang, and Ming- sheng Long. 2024. Simmtm: A simple pre-training framework for masked time- series modeling.AdvancesinNeuralInformationProcessingSystems36 (2024). [4]Chiming Duan, Minghua He, Pei Xiao, Tong Jia, Xin Zhang, Zhewei Zhong, Xiang Luo, Yan Niu, Lingzhe Zhang, Yifan Wu, et al.2025. LogAction: Consistent Cross- system Anomaly Detection through Logs via Active Domain.arXivpreprint arXiv:2510.03288 (2025). [5]Stephen Elliot. 2014. DevOps and the cost of downtime: Fortune 1000 best practice metrics quantified.InternationalDataCorporation(IDC) (2014). [6]Yu Gan, Mingyu Liang, Sundar Dev, David Lo, and Christina Delimitrou. 2021. Sage: practical and scalable ML-driven performance debugging in microser- vices. InProceedingsofthe26thACMInternationalConferenceonArchitectural SupportforProgrammingLanguagesandOperatingSystems. 135–151. [7] Google Cloud Platform. 2025. Online Boutique: A Cloud-First Microservices Demo Application. https://github.com/GoogleCloudPlatform/microservices- demo. Accessed: October 15, 2025. [8]Minghua He, Chiming Duan, Pei Xiao, Tong Jia, Siyu Yu, Lingzhe Zhang, Weijie Hong, Jin Han, Yifan Wu, Ying Li, et al.2025. United we stand: Towards end-to- end log-based fault diagnosis via interactive multi-task learning.arXivpreprint arXiv:2509.24364 (2025). [9]Minghua He, Tong Jia, Chiming Duan, Pei Xiao, Lingzhe Zhang, Kangjin Wang, Yi- fan Wu, Ying Li, and Gang Huang. 2025. Walk the talk: Is your log-based software reliability maintenance system really reliable?arXivpreprintarXiv:2509.24352 (2025). [10] Weijie Hong, Yifan Wu, Lingzhe Zhang, Chiming Duan, Pei Xiao, Minghua He, Xixuan Yang, and Ying Li. 2025. CSLParser: A Collaborative Framework Using Small and Large Language Models for Log Parsing. In2025IEEE36th InternationalSymposiumonSoftwareReliabilityEngineering(ISSRE). IEEE, 61– 72. [11]Xiaosong Huang, Hongyi Liu, Yifan Wu, Lingzhe Zhang, Tong Jia, Ying Li, and Zhonghai Wu. 2025. UDA-RCL: Unsupervised Domain Adaptation for Microser- vice Root Cause Localization Utilizing Multimodal Data.IEEETransactionson ServicesComputing (2025). [12] Information Technology Intelligence Consulting (ITIC). 2024.ITIC2024Global ServerHardware,ServerOSReliabilityReport. Annual Report. ITIC. [13]Cheryl Lee, Tianyi Yang, Zhuangbin Chen, Yuxin Su, and Michael R Lyu. 2023. Eadro: An end-to-end troubleshooting framework for microservices on multi- source data. In2023IEEE/ACM45thInternationalConferenceonSoftware Engineering(ICSE). IEEE, 1750–1762. [14]Xiaoyun Li, Pengfei Chen, Linxiao Jing, Zilong He, and Guangba Yu. 2022. Swiss- Log: Robust anomaly detection and localization for interleaved unstructured logs.IEEETransactionsonDependableandSecureComputing20, 4 (2022), 2762–2780. [15] Yufeng Li, Guangba Yu, Pengfei Chen, Chuanfu Zhang, and Zibin Zheng. 2022. MicroSketch: Lightweight and adaptive sketch based performance issue detec- tion and localization in microservice systems. InInternationalConferenceon Service-OrientedComputing. Springer, 219–236. [16] Wenlong Liao, Shouxiang Wang, Dechang Yang, Zhe Yang, Jiannong Fang, Chris- tian Rehtanz, and Fernando Porté-Agel. 2025. TimeGPT in load forecasting: A large time series model perspective.AppliedEnergy 379 (2025), 124973. [17] Cheng-Ming Lin, Ching Chang, Wei-Yao Wang, Kuang-Da Wang, and Wen-Chih Peng. 2024. Root Cause Analysis in Microservice Using Neural Granger Causal Discovery. InProceedingsoftheAAAIConferenceonArtificialIntelligence, Vol. 38. 206–213. [18]Hongyi Liu, Xiaosong Huang, Mengxi Jia, Tong Jia, Jing Han, Zhonghai Wu, and Ying Li. 2024. Uac-ad: Unsupervised adversarial contrastive learning for anomaly detection on multi-modal data in microservice systems.IEEETransactionson ServicesComputing 17, 6 (2024), 3887–3900. [19]Hongyi Liu, Xiaosong Huang, Mengxi Jia, Lingzhe Zhang, Tong Jia, Zhonghai Wu, and Ying Li. 2025. AAAD: Asynchronous Inter-Variable Relationship-Aware Anomaly Detection for Multivariate Time Series. In2025IEEEInternational ConferenceonMultimediaandExpo(ICME). IEEE, 1–6. [20] Hongyi Liu, Yinping Ma, Xiaosong Huang, Lingzhe Zhang, Tong Jia, and Ying Li. 2025. ORA: Job Runtime Prediction for High-Performance Computing Platforms Using the Online Retrieval-Augmented Language Model. InProceedingsofthe 39thACMInternationalConferenceonSupercomputing. 884–894. [21]Kashif Rasul, Arjun Ashok, Andrew Robert Williams, Arian Khorasani, George Adamopoulos, Rishika Bhagwatkar, Marin Biloš, Hena Ghonia, Nadhir Hassen, Anderson Schneider, et al.2023. Lag-llama: Towards foundation models for time series forecasting. InR0-FoMo:RobustnessofFew-shotandZero-shotLearning inLargeFoundationModels. [22]Yicheng Sui, Yuzhe Zhang, Jianjun Sun, Ting Xu, Shenglin Zhang, Zhengdan Li, Yongqian Sun, Fangrui Guo, Junyu Shen, Yuzhi Zhang, et al.2023. Logkg: Log failure diagnosis through knowledge graph.IEEETransactionsonServices Computing 16, 5 (2023), 3493–3507. [23] Yongqian Sun, Zihan Lin, Binpeng Shi, Shenglin Zhang, Shiyu Ma, Pengxiang Jin, Zhenyu Zhong, Lemeng Pan, Yicheng Guo, and Dan Pei. 2025. Interpretable failure localization for microservice systems based on graph autoencoder.ACM TransactionsonSoftwareEngineeringandMethodology 34, 2 (2025), 1–28. [24] Yongqian Sun, Binpeng Shi, Mingyu Mao, Minghua Ma, Sibo Xia, Shenglin Zhang, and Dan Pei. 2024. Art: A unified unsupervised framework for incident management in microservice systems. InProceedingsofthe39thIEEE/ACM InternationalConferenceonAutomatedSoftwareEngineering. 1183–1194. Conference’17, July 2017, Washington, DC, USALingzhe Zhang et al. [25]Guangba Yu, Pengfei Chen, Hongyang Chen, Zijie Guan, Zicheng Huang, Linxiao Jing, Tianjun Weng, Xinmeng Sun, and Xiaoyun Li. 2021. Microrank: End-to- end latency issue localization with extended spectrum analysis in microservice environments. InProceedingsoftheWebConference2021. 3087–3098. [26]Guangba Yu, Pengfei Chen, Yufeng Li, Hongyang Chen, Xiaoyun Li, and Zibin Zheng. 2023. Nezha: Interpretable fine-grained root causes analysis for microservices on multi-modal observability data. InProceedingsofthe31st ACMJointEuropeanSoftwareEngineeringConferenceandSymposiumonthe FoundationsofSoftwareEngineering. 553–565. [27]Guangba Yu, Zicheng Huang, and Pengfei Chen. 2023. TraceRank: Abnormal service localization with dis-aggregated end-to-end tracing data in cloud native systems.JournalofSoftware:EvolutionandProcess 35, 10 (2023), e2413. [28] Chenxi Zhang, Zhen Dong, Xin Peng, Bicheng Zhang, and Miao Chen. 2024. Trace-based Multi-Dimensional Root Cause Localization of Performance Issues in Microservice Systems. InProceedingsoftheIEEE/ACM46thInternational ConferenceonSoftwareEngineering. 1–12. [29]Lingzhe Zhang, Tong Jia, Mengxi Jia, Ying Li, Yong Yang, and Zhonghai Wu. 2024. Multivariate log-based anomaly detection for distributed database. In Proceedingsofthe30thACMSIGKDDConferenceonKnowledgeDiscovery andDataMining. 4256–4267. [30]Lingzhe Zhang, Tong Jia, Mengxi Jia, Yifan Wu, Aiwei Liu, Yong Yang, Zhonghai Wu, Xuming Hu, Philip Yu, and Ying Li. 2025. A Survey of AIOps in the Era of Large Language Models.Comput.Surveys (2025). [31]Lingzhe Zhang, Tong Jia, Mengxi Jia, Yifan Wu, Hongyi Liu, and Ying Li. 2025. ScalaLog: Scalable Log-Based Failure Diagnosis Using LLM. InICASSP2025-2025 IEEEInternationalConferenceonAcoustics,SpeechandSignalProcessing (ICASSP). IEEE, 1–5. [32] Lingzhe Zhang, Tong Jia, Mengxi Jia, Yifan Wu, Hongyi Liu, and Ying Li. 2025. XRAGLog: A Resource-Efficient and Context-Aware Log-Based Anomaly Detec- tion Method Using Retrieval-Augmented Generation. InAAAI2025Workshop onPreventingandDetectingLLMMisinformation(PDLM). [33] Lingzhe Zhang, Tong Jia, Xinyu Tan, Xiangdong Huang, Mengxi Jia, Hongyi Liu, Zhonghai Wu, and Ying Li. 2025. E-log: Fine-grained elastic log-based anomaly detection and diagnosis for databases.IEEETransactionsonServicesComputing (2025). [34] Lingzhe Zhang, Tong Jia, Kangjin Wang, Weijie Hong, Chiming Duan, Minghua He, and Ying Li. 2025. Adaptive root cause localization for microservice systems with multi-agent recursion-of-thought.arXivpreprintarXiv:2508.20370(2025). [35]Lingzhe Zhang, Tong Jia, Kangjin Wang, Mengxi Jia, Yong Yang, and Ying Li. 2024. Reducing events to augment log-based anomaly detection models: An empirical study. InProceedingsofthe18thACM/IEEEInternationalSymposium onEmpiricalSoftwareEngineeringandMeasurement. 538–548. [36]Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Leyi Pan, Chiming Duan, Minghua He, Mengxi Jia, and Ying Li. 2026. Agentic Memory Enhanced Recursive Reasoning for Root Cause Localization in Microservices.arXivpreprintarXiv:2601.02732 (2026). [37] Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Leyi Pan, Chiming Duan, Minghua He, Pei Xiao, and Ying Li. 2026. Hypothesize-Then-Verify: Speculative Root Cause Analysis for Microservices with Pathwise Parallelism.arXivpreprint arXiv:2601.02736 (2026). [38]Lingzhe Zhang, Yunpeng Zhai, Tong Jia, Chiming Duan, Minghua He, Leyi Pan, Zhaoyang Liu, Bolin Ding, and Ying Li. 2025. MicroRemed: Benchmarking LLMs in Microservices Remediation.arXivpreprintarXiv:2511.01166 (2025). [39]Lingzhe Zhang, Yunpeng Zhai, Tong Jia, Chiming Duan, Siyu Yu, Jinyang Gao, Bolin Ding, Zhonghai Wu, and Ying Li. 2025. ThinkFL: Self-Refining Failure Localization for Microservice Systems via Reinforcement Fine-Tuning.arXiv preprintarXiv:2504.18776 (2025). [40]Lingzhe Zhang, Yunpeng Zhai, Tong Jia, Xiaosong Huang, Chiming Duan, and Ying Li. 2025. Agentfm: Role-aware failure management for distributed databases with llm-driven multi-agents. InProceedingsofthe33rdACMInternational ConferenceontheFoundationsofSoftwareEngineering. 525–529. [41]Shenglin Zhang, Pengxiang Jin, Zihan Lin, Yongqian Sun, Bicheng Zhang, Sibo Xia, Zhengdan Li, Zhenyu Zhong, Minghua Ma, Wa Jin, et al.2023. Robust failure diagnosis of microservice system through multimodal data.IEEETransactions onServicesComputing 16, 6 (2023), 3851–3864. [42]Zhizhou Zhang, Murali Krishna Ramanathan, Prithvi Raj, Abhishek Parwal, Timothy Sherwood, and Milind Chabbi. 2022. CRISP: Critical path analysis of Large-Scale microservice architectures. In2022USENIXAnnualTechnical Conference(USENIXATC22). 655–672. [43]Xiang Zhou, Xin Peng, Tao Xie, Jun Sun, Chao Ji, Wenhai Li, and Dan Ding. 2018. Fault analysis and debugging of microservice systems: Industrial survey, bench- mark system, and empirical study.IEEETransactionsonSoftwareEngineering 47, 2 (2018), 243–260.