Zexin Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Zexin Wang has a total of 39 co-authors. Here are the top co-authors by number of shared publications:
- Changhua Pei
Computer Network Information Center, Chinese Academy of Sciences, Beijing, China
10 shared publications - Dan Pei
Tsinghua University, Beijing, China
10 shared publications - Gaogang Xie
Chinese Academy of Sciences
10 shared publications - Jingjing Li
Computer Network Information Center, Chinese Academy of Sciences, Beijing, China
8 shared publications - Hang Cui
Computer Network Information Center, Chinese Academy of Sciences, Beijing, China
6 shared publications - Cenjie Hu
Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang, China
5 shared publications
Latest Publications
- Open Access
- 7 October 2026
- Published by Association for Computing Machinery (ACM)
- p. 1767-1779
Smart Brain: Semantic Anomaly Detection for Operational Time Series in Large Scale Service Systems
- 7 October 2026
- Published by Association for Computing Machinery (ACM)
- p. 1767-1779
Open AccessSmart Brain: Semantic Anomaly Detection for Operational Time Series in Large Scale Service Systems
- Open Access
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 2253-2255
Predict Boldly, Recover Cautiously: Fast On-Router Route Anomaly Prediction and Recovery
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 2253-2255
Open AccessPredict Boldly, Recover Cautiously: Fast On-Router Route Anomaly Prediction and Recovery
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 710-720
Rethinking Time Series Anomaly Detection from a Dynamic Perspective: Temporal-Frequency-Curvature Fusion
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 710-720
Open AccessRethinking Time Series Anomaly Detection from a Dynamic Perspective: Temporal-Frequency-Curvature Fusion
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 7773-7784
Don't Predict, Prioritize: Rethinking GPU Reliability Assessment
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 7773-7784
Open AccessDon't Predict, Prioritize: Rethinking GPU Reliability Assessment
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 8575-8584
TSLoc: Self-Supervised Faulty Node Localization Framework in Large-scale Training Clusters
As model parameters and GPU cluster sizes scale, detecting faulty nodes becomes increasingly complex, where a single failure can idle thousands of GPUs. Recent work has accelerated fault localization by comparing metric time-series, but these approaches are often brittle. They typically rely on decision trees trained on historical failure data to prioritize metrics, which limits generalization to unseen failures and prevents day-one support for new monitoring metrics. To address this brittleness, we propose TSLoc, a resilient and general fault localization framework. TSLoc introduces a novel unified time-series model that learns a universal representation of normal synchronous behavior across diverse metrics, identifying faulty nodes by their deviation from this learned norm without needing prior fault signatures. In our evaluation on real-world failures, TSLoc not only provides broad coverage for common faults, achieving a Top-5 accuracy of 0.908 with an average processing time of 4.2 seconds, but also proves highly effective on challenging long-tail faults where prior methods struggle.
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 8575-8584
Open AccessTSLoc: Self-Supervised Faulty Node Localization Framework in Large-scale Training Clusters
As model parameters and GPU cluster sizes scale, detecting faulty nodes becomes increasingly complex, where a single failure can idle thousands of GPUs. Recent work has accelerated fault localization by comparing metric time-series, but these approaches are often brittle. They typically rely on decision trees trained on historical failure data to prioritize metrics, which limits generalization to unseen failures and prevents day-one support for new monitoring metrics. To address this brittleness, we propose TSLoc, a resilient and general fault localization framework. TSLoc introduces a novel unified time-series model that learns a universal representation of normal synchronous behavior across diverse metrics, identifying faulty nodes by their deviation from this learned norm without needing prior fault signatures. In our evaluation on real-world failures, TSLoc not only provides broad coverage for common faults, achieving a Top-5 accuracy of 0.908 with an average processing time of 4.2 seconds, but also proves highly effective on challenging long-tail faults where prior methods struggle.
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 9071-9081
TSRBench: Benchmarking Time-Series Retrieval
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 9071-9081
Open AccessTSRBench: Benchmarking Time-Series Retrieval
- Open Access
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 5286-5297
ViTs: Teaching Machines to See Time Series Anomalies Like Human Experts
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 5286-5297
Open AccessViTs: Teaching Machines to See Time Series Anomalies Like Human Experts
- Open Access
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 7869-7879
Smart Eye: LLM-Guided Proposer-Verifier Framework for Industrial-Scale Log Anomaly Detection
A practical log anomaly detection system for large-scale web services, suitable for real industrial deployment, should exhibit three key characteristics: fidelity to decisive lexical cues, human-auditable explanations, and low inference latency. Existing approaches often compromise one or more of these requirements by discarding crucial tokens during parsing, obscuring decisions through dense embeddings, or incurring heavy computational costs. We introduce Smart Eye, a deployable two-stage key pattern retrieval framework that employs large language models (LLMs) strictly as proposal engines, coupled with a deterministic verifier. In Stage I, the system enhances recall of anomalous logs under a configurable false positive budget using budgeted maximum coverage selection. In Stage II, precision is monotonically improved by refining parent patterns into specialized child patterns and introducing targeted exclusions. These essential patterns, extracted by LLMs and then confirmed, will be formulated into regular expressions to enable efficient online matching of anomalous logs. We provide both theoretical justification and a practical implementation that runs efficiently on commodity CPUs. When evaluated on three real world Huawei Cloud Core log datasets, Smart Eye achieves state-of-the-art anomaly detection with a average F1 score of 0.9877, outperforming the strongest neural baseline. Moreover, it produces human interpretable evidence chains and supports low latency inference. These results demonstrate that combining LLM-guided proposal generation with deterministic verification forms a robust and deployable alternative to traditional end to end embedding pipelines for industrial log analysis.
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 7869-7879
Open AccessSmart Eye: LLM-Guided Proposer-Verifier Framework for Industrial-Scale Log Anomaly Detection
A practical log anomaly detection system for large-scale web services, suitable for real industrial deployment, should exhibit three key characteristics: fidelity to decisive lexical cues, human-auditable explanations, and low inference latency. Existing approaches often compromise one or more of these requirements by discarding crucial tokens during parsing, obscuring decisions through dense embeddings, or incurring heavy computational costs. We introduce Smart Eye, a deployable two-stage key pattern retrieval framework that employs large language models (LLMs) strictly as proposal engines, coupled with a deterministic verifier. In Stage I, the system enhances recall of anomalous logs under a configurable false positive budget using budgeted maximum coverage selection. In Stage II, precision is monotonically improved by refining parent patterns into specialized child patterns and introducing targeted exclusions. These essential patterns, extracted by LLMs and then confirmed, will be formulated into regular expressions to enable efficient online matching of anomalous logs. We provide both theoretical justification and a practical implementation that runs efficiently on commodity CPUs. When evaluated on three real world Huawei Cloud Core log datasets, Smart Eye achieves state-of-the-art anomaly detection with a average F1 score of 0.9877, outperforming the strongest neural baseline. Moreover, it produces human interpretable evidence chains and supports low latency inference. These results demonstrate that combining LLM-guided proposal generation with deterministic verification forms a robust and deployable alternative to traditional end to end embedding pipelines for industrial log analysis.
- Top Cited PapersOpen Access
- 8 May 2025
- Published by Association for Computing Machinery (ACM)
- p. 422-431
Flow-of-Action: SOP Enhanced LLM-Based Multi-Agent System for Root Cause Analysis
- 8 May 2025
- Published by Association for Computing Machinery (ACM)
- p. 422-431
Top Cited PapersOpen AccessFlow-of-Action: SOP Enhanced LLM-Based Multi-Agent System for Root Cause Analysis
- Open Access
- 13 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 3096-3105
Revisiting VAE for Unsupervised Time Series Anomaly Detection: A Frequency Perspective
Time series Anomaly Detection (AD) plays a crucial role for web systems. Various web systems rely on time series data to monitor and identify anomalies in real time, as well as to initiate diagnosis and remediation procedures. Variational Autoencoders (VAEs) have gained popularity in recent decades due to their superior de-noising capabilities, which are useful for anomaly detection. However, our study reveals that VAE-based methods face challenges in capturing long-periodic heterogeneous patterns and detailed short-periodic trends simultaneously. To address these challenges, we propose Frequency-enhanced Conditional Variational Autoencoder (FCVAE), a novel unsupervised AD method for univariate time series. To ensure an accurate AD, FCVAE exploits an innovative approach to concurrently integrate both the global and local frequency features into the condition of Conditional Variational Autoencoder (CVAE) to significantly increase the accuracy of reconstructing the normal data. Together with a carefully designed "target attention" mechanism, our approach allows the model to pick the most useful information from the frequency domain for better short-periodic trend construction. Our FCVAE has been evaluated on public datasets and a large-scale cloud system, and the results demonstrate that it outperforms state-of-the-art methods. This confirms the practical applicability of our approach in addressing the limitations of current VAE-based anomaly detection models.
- 13 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 3096-3105
Open AccessRevisiting VAE for Unsupervised Time Series Anomaly Detection: A Frequency Perspective
Time series Anomaly Detection (AD) plays a crucial role for web systems. Various web systems rely on time series data to monitor and identify anomalies in real time, as well as to initiate diagnosis and remediation procedures. Variational Autoencoders (VAEs) have gained popularity in recent decades due to their superior de-noising capabilities, which are useful for anomaly detection. However, our study reveals that VAE-based methods face challenges in capturing long-periodic heterogeneous patterns and detailed short-periodic trends simultaneously. To address these challenges, we propose Frequency-enhanced Conditional Variational Autoencoder (FCVAE), a novel unsupervised AD method for univariate time series. To ensure an accurate AD, FCVAE exploits an innovative approach to concurrently integrate both the global and local frequency features into the condition of Conditional Variational Autoencoder (CVAE) to significantly increase the accuracy of reconstructing the normal data. Together with a carefully designed "target attention" mechanism, our approach allows the model to pick the most useful information from the frequency domain for better short-periodic trend construction. Our FCVAE has been evaluated on public datasets and a large-scale cloud system, and the results demonstrate that it outperforms state-of-the-art methods. This confirms the practical applicability of our approach in addressing the limitations of current VAE-based anomaly detection models.