Guanqun Yang 杨冠群
 |
Ph.D. Candidate in Computer Science Email: gyang16@stevens.edu or guanqun.yang@outlook.com Office: Room 433, Gateway South Hall (Google Map) LinkedIn | GitHub | Google Scholar | Semantic Scholar | ORCID |
About Me
My research focuses on scalable software testing and analysis, with recent emphasis on testing and securing AI systems including LLMs and multi-agent architectures. I develop automated solutions that reduce manual effort in ML model validation, security vulnerability management, and evaluating AI coding assistants.
AI System Testing: Developed automated test generation frameworks (TestAug, HateModerate) that reduce manual annotation by 98% while uncovering failures in production ML models. Currently building user simulation frameworks to evaluate AI coding assistants.
Multi-Agent Systems and Security: Building and securing agentic workflows using LangGraph, MCP, and agent-to-agent protocols. Discovered prompt-injection vulnerabilities in MCP tool schemas affecting LLM agent behavior.
LLM Fine-Tuning and Alignment: Extensive experience with SFT, RLHF, and GRPO for model customization. Fine-tuned LLMs to reduce hallucination, control verbosity, and generate constrained outputs for red-teaming.
Retrieval-Augmented Generation: Developed scalable retrieval systems using vector databases (FAISS, ChromaDB) and embedding models for security patch retrieval and knowledge-intensive applications.
I am fortunate to be advised by Prof. Xueqing (Susan) Liu. Before coming to Stevens, I received my master's degree in Electrical and Computer Engineering from UCLA in 2019. I was working with Prof. Quanquan Gu and Prof. Vwani P. Roychowdhury on algorithmic fairness in machine learning systems (my master thesis). I received my Bachelor's degree in Electrical Engineering from Northeastern University, China in 2017.
Education
Industry Experience
Capital One, McLean, VA
Data Scientist Intern – LLM-based Legal Validation (OpenReview) (Summer 2025)
Built an explainable multi-LLM debate system to validate the legality of machine-generated marketing leads. Customized baseline models via SFT and GRPO fine-tuning on AWS EKS, and integrated RAG for context awareness, achieving over 10% improvements in accuracy and F1.
Data Scientist Intern – Hallucination and Verbosity Control for RAG-LLM Chatbots (Summer 2024)
Built a data synthesis pipeline that transforms heterogeneous call center and site content logs into QA-style SFT data. Reduced verbosity by 89.67% and hallucination by 201.68% by fine-tuning the LLM on curated synthetic SFT data.
Technical Skills
Deep Learning and LLMs
Frameworks: PyTorch, Hugging Face Transformers, DeepSpeed
LLM Training: SFT, RLHF, GRPO, parameter-efficient fine-tuning (LoRA, QLoRA)
Inference: vLLM
Agentic AI and RAG
Agent Frameworks: LangGraph, LangChain, AutoGen (AG2), Google ADK, smolagents
Protocols: MCP, A2A Protocol
Vector Databases: FAISS, ChromaDB, Milvus
Knowledge Graphs: Neo4j, GraphRAG
Infrastructure and Cloud
Cloud: AWS (EKS, EC2), GCP
MLOps: Docker, Weights & Biases
Data: ElasticSearch
Recent Activities
[2025-08] Researching security vulnerabilities in agentic LLM workflows (MCP, LangGraph).
[2025-06] Started second internship at Capital One, building LLM-based legal validation systems.
[2025-05] Preprint on scalable security patch retrieval released on arXiv.
[2024-06] Started interning at Capital One, fine-tuning LLMs for RAG chatbots.
[2024-02] Paper accepted at NAACL 2024 Findings: HateModerate (WOAH 2024 Outstanding Paper).
[2023-07] Gave a tutorial on AI and ChatGPT to underrepresented high school students. [Course Materials]