School of Engineering Department of Computer Science and Engineering 152 Open World Understanding Based on Large Vision-Language Models Supervisor: XU Dan / CSE Student: KWOK Yuk Hang / COMP Course: UROP 1100, Fall Open-world understanding aims to leverage large sets of object classes to provide a semantic understanding of the real world, which contains unpredictable, never-seen situations. Performance in zero-shot settings, which seeks to solve tasks that are not observed during training, is one of the frontiers in machine learning. In most cases, we deal with multiple unseen objects simultaneously. For a while, RCNN and FCN architectures have been two common ways of performing image understanding. In recent years, Vision-Language Pretraining (VLP) models have gained attention with the promising open-vocabulary results from CLIP and ALIGN architecture. In this paper, we explore various strategies that make use of large vision-language models to perform visual perception and open-world understanding of the real world. A LLM-Based Log Parsing, Variable Annotation, and Anomaly Detection System Supervisor: YI Ke / CSE Student: DING Shuye / MATH-CS Course: UROP 1100, Summer This literature review examines LLM-based methods for extracting fault-indicating information from logs in failure diagnosis, focusing on five representative works employing prompt engineering, chain-of-thought (CoT) reasoning, in-context learning (ICL), and semantic-aware parsing. These studies address tasks like faultindicating description (FID) and parameter (FIP) extraction, anomaly detection, and log summarization, evaluated in both open-source and industrial datasets. Key advancements include LoFI for prompt-based tuning on PLMs, LogGPT for zero-shot log analysis with LLMs, and PromptLogParser for efficient log parsing via prompts. The review compares methodologies, applications, and performance metrics, highlighting trade-offs in accuracy, data efficiency, and scalability. Experimental results from prompt variations show precision improvements up to 0.588 for FID and 0.466 for FIP on the FIBench dataset, outperforming baseline LLMs but trailing specialized models like LoFI. Results highlight the potential and limitations of generalpurpose LLMs for log analysis. A LLM-Based Log Parsing, Variable Annotation, and Anomaly Detection System Supervisor: YI Ke / CSE Student: LIU Teng / CPEG Course: UROP 1100, Summer This literature review surveys the application of large language models (LLMs) in log parsing, focusing on employing In-Context Learning (ICL) to enhance log parsing capabilities and improving efficiency through methods such as self-caching. These studies tackle issues like domain terminology gaps, template variability in logs from heterogeneous systems, and the low efficiency of LLMs. This review is based on two papers: DivLog and LILAC. Key innovations include DivLog’s prompt-enhanced ICL for training-free template extraction and LILAC’s adaptive parsing cache for efficient and consistent parsing. Furthermore, we also conducted experiments to test and improve the DivLog system, with the aim of enabling a better understanding and utilization of the log parsing methods of DivLog.
RkJQdWJsaXNoZXIy NDk5Njg=