Zegang Peng

Zegang Peng

Auto-generated profile, based on publicly available information

Zegang Peng

Tsinghua University, Beijing, China

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
30 June 2026 in Proceedings of the ACM on Software Engineering
Verified
Journal Article
Open access
JMJiayi Mao
Jiayi Mao
LLLiqun Li
Liqun Li
YGYanjie Gao
Yanjie Gao
ZPZegang Peng
Zegang Peng
SHShilin He
Shilin He
DZDongmei Zhang
Dongmei Zhang
Effective incident management in large-scale IT systems relies on troubleshooting guides (TSGs), but their manual execution is slow and error-prone. While recent advances in LLMs offer promise for automating incident management tasks, existing LLM-based solutions lack specialized support for several key challenges, including managing TSG quality issues, interpreting complex control flow, handling data-intensive queries, and exploiting execution parallelism. We first conducted an empirical study on 92 real-world TSGs, and, guided by our findings, we present , a novel end-to-end agentic framework for troubleshooting guide automation. Our approach features a three-stage workflow: the first stage provides a comprehensive guide together with a tool, TSG Mentor, to assist site reliability engineers (SREs) in improving TSG quality; the second stage performs offline preprocessing using LLMs to extract structured execution directed acyclic graphs (DAGs) from unstructured TSGs and to create dedicated Query Preparation Plugins (QPPs); and the third stage executes online using a DAG-guided scheduler-executor framework with a memory system to ensure correct workflow and support parallel execution of independent steps. Our empirical evaluation on a collection of real-world TSGs and incidents demonstrates that achieves a ∼94% success rate on GPT-4.1, outperforming baselines with less time and token consumption. Furthermore, it achieves a remarkable execution time reduction of 32.9% to 70.4% for parallelizable TSGs. Our code and sample data are publicly available at https://github.com/microsoft/StepFly.

1 January 2026 in IEEE Transactions on Image Processing
Verified
Journal Article
Closed access
HZHanle Zheng
Hanle Zheng
XHXujie Han
Xujie Han
ZPZegang Peng
Zegang Peng
SZShangbin Zhang
Shangbin Zhang
GDGuangxun Du
Guangxun Du
Lei DengLei Deng
Lei Deng

18 July 2025 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Verified
Journal Article
Closed access
XQXiaojun Qi
Xiaojun Qi
YYYujie Ying
Yujie Ying
ZZZhong Zheng
Zhong Zheng
ZYZiyi Yang
Ziyi Yang
LLLing Liang
Ling Liang
Lei DengLei Deng
Lei Deng
Hybrid brain-inspired computing (BIC), which integrates computer science-oriented artificial neural networks (ANNs) with neuroscience-oriented spiking neural networks (SNNs), represents a significant trend in the field of BIC. This approach demonstrates considerable advantages in domains, such as perception, cognition, and learning. Nevertheless, several challenges persist in deploying hybrid BIC on resource-constrained edge devices, particularly in enabling these devices to effectively manage complex tasks. To tackle these challenges, we present eBrainISA, an edge-oriented hybrid BIC instruction set architecture (ISA). eBrainISA is designed to support sophisticated ANNs, SNNs, and hybrid neural networks (HNNs), which integrates an extensive suite of neural operators that are tailored to accommodate the hardware limitations on edge devices. To validate the proposed architecture, we develop a unified low-precision quantization approach alongside a hardware simulator. Prior ISAs offer limited support for hybrid BIC, while ours can support all major networks, including ANNs such as ResNet and Yolo, SNNs such as Spike_ResNet and Spike_Transformer, and HNNs such as DetectHNN and ReSpike with <1% accuracy degradation. Moreover, our architecture achieves about $5\times $ memory saving compared to full-precision versions and $3.87\times $ – $53.61\times $ higher energy efficiency compared to edge GPU when executing different neural networks with diverse behaviors. This work demonstrates a promising balance between intelligent demands for general models and resource constraints on edge devices.