Universiti Sains Malaysia (USM)

Master of Computer Science

Oct 2023 – Mar 2025

GPA: 3.3 / 4.0

Relevant Courses:

• Deep Learning
• Multimodal Learning
• Internet of Things
• Algorithm Design

Yingkou Institute of Technology

Bachelor of Intelligent Science and Technology

Sep 2018 – Jun 2022

GPA: 3.2 / 4.0

Honors:

• Outstanding Graduate

Relevant Courses:

• Machine Learning
• Data Mining
• Big Data Analytics
• Data Visualization

360 Digital Security Group

Large Model Algorithm Engineer

Aug 2025 – Present

WatchLog (EDR Log Security Analysis Framework) ICML 2026 Regular

• Proposed and developed WatchLog, a multimodal LLM framework that transforms ultra-long EDR log sequences (up to 1M tokens) into video-like structures for efficient and interpretable event reasoning.
• Designed a novel kvEmbedding layer to encode key-value log events into image tensors, and applied temporal cross-attention to compress long sequences into fixed-length representations, significantly reducing computational overhead.
• Implemented a three-stage training pipeline (image-event alignment, video-log alignment, and supervised fine-tuning) to simultaneously output threat categories and natural language reasoning traces.
• Achieved 99.8% binary classification accuracy and 100% recall on the custom EDR8M-20R dataset, outperforming traditional methods and fine-tuned LLMs.
• Reduced inference latency by over 200x (first-token latency of 1.11s) and GPU memory usage by 50.6% (48.6GB for 1024K tokens) compared to Qwen2.5-7B, enabling single-GPU processing of million-token logs.

PSIFrame (Self-Evolving LLM Multi-Agent Framework) NeurIPS 2026 Under Review

• Developed PSIFrame, a self-evolving multi-agent framework designed to overcome the limitations of static workflows and flat tool selection in dynamic, open-world environments.
• Constructed a hierarchical Taxonomy Graph for tool organization, enabling top-down, task-aware routing through functional and executable agents.
• Integrated Monte Carlo Tree Search (MCTS) with an adversarial debate module for adaptive planning, execution, and dynamic backtracking based on environmental feedback.
• Improved tool retrieval NDCG@10 by 21.3% and Recall@10 by 19.8% on the MetaTool dataset, while reducing token consumption by ~50%.
• Achieved an 80.0% task success rate on a custom open-world multi-agent benchmark (BenchVerse), significantly outperforming baselines like MetaAgent (68.0%).

NDR Large Model Distillation (Ongoing)

• Conducting data analysis on NDR log JSON files and fine-tuning Qwen3-14B using LlamaFactory with a CoE architecture.
• Optimizing algorithms to address long-tail distribution problems, successfully improving model performance from 82% to 93% to date.

SF Technology

AI Engineer Intern

Feb 2025 – Jun 2025

Project

Multimodal Financial Shared-Service LLM System

Technologies

• Qwen
• LoRA
• SFT
• OCR
• YOLOv5
• PyTorch

Contributions

• Fine-tuned multimodal LLMs (e.g., Qwen-VL, LLaVA) using LoRA and SFT for logistics and financial document understanding, improving information extraction accuracy from 85% to 96%.
• Built and optimized OCR pipelines using YOLOv5 and CRNN for complex invoice and receipt parsing, increasing parsing precision by 15% and reducing processing time per document by 40%.
• Applied object detection models for automated cargo verification and anomaly detection, enhancing the robust performance of the vision system under varying lighting conditions.
• Integrated multimodal reasoning capabilities into financial shared-service workflows, reducing manual verification workload by 60%.

Lazada

Data Mining & Recommendation Intern

Jun 2024 – Dec 2024

Technologies

• PCA
• LDA
• SVM
• KNN
• XGBoost
• Collaborative Filtering

Contributions

• Conducted deep-dive CTR and CVR analysis on user behavior data to identify key features driving user engagement and conversion.
• Built and deployed robust recommendation and ranking models using XGBoost and LightGBM, improving overall CTR by 12% and CVR by 8%.
• Implemented collaborative filtering algorithms (UserCF, ItemCF) and matrix factorization techniques (SVD, NMF) to alleviate the cold-start problem, increasing recommendation coverage by 20%.
• Applied Learning-to-Rank (L2R) techniques to optimize the ranking phase of the recommendation pipeline, yielding a 5% improvement in NDCG@10.

Hong Kong University of Science and Technology (Guangzhou)

Research Assistant

Jan 2025 – Aug 2025

Topic

Replay-Free Continual Graph Learning

Research Problem

Graph neural networks suffer from catastrophic forgetting when new classes or tasks arrive continuously.

Contributions

• Designed an analytic learning framework for continual graph learning.
• Eliminated replay buffers to preserve privacy and reduce storage cost.
• Developed recursive least squares based memory updates.
• Constructed recursive correlation memory matrices for stable knowledge retention.
• Improved training efficiency while maintaining competitive performance.

Representative Work

AL-GNN: Privacy-Preserving and Replay-Free Continual Graph Learning via Analytic Learning

Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences

Research Collaborator

Dec 2024 – Jun 2025

Topic

Multimodal Large Language Model Alignment

Research Problem

Existing multimodal foundation models show weak generalization on scientific and chemical reasoning tasks.

Contributions

• Built a benchmark covering eight multimodal chemistry tasks.
• Investigated DPO and GRPO based multimodal alignment strategies.
• Improved cross-modal consistency through balanced optimization.
• Enhanced transferability across heterogeneous reasoning tasks.