LLM Agent Safety
Locating and attributing failures in LLM agent reasoning during multi-step, safety-critical tasks, and compiling verified failures into runtime checks.
Shou-Tzu Han (Debra Han)
Computer Science Ph.D. student and Graduate Research Assistant at Wayne State University.
01
Research Profile
Computer Science Ph.D. student and Graduate Research Assistant at Wayne State University specializing in trustworthy AI, LLM agent safety, LLM robustness, and mechanistic interpretability. Experienced in designing reproducible GPU/HPC experiments, analyzing internal model behavior, and building evaluation pipelines for language and agent systems. Author of three arXiv preprints. Seeking a research internship focused on reliable and interpretable AI systems.
02
Research Interests
LLM Robustness and Reasoning
Evaluating reasoning stability under meaning-preserving perturbations and diagnosing performance-confidence mismatches.
Mechanistic Interpretability
Using activation analysis, patching, and ablation to identify internal components associated with model failures.
Agentic Retrieval and Scientific Discovery
Building multi-step retrieval systems for literature comparison, contribution analysis, and research-gap identification.
03
Publications
012026
Fragile Reasoning: A Mechanistic Analysis of LLM Sensitivity to Meaning-Preserving Perturbations
Preprint · arXiv:2604.01639
022026
Novelty-Aware Agentic Retrieval: Comparing Research Contributions Through Structured Multi-Step Reasoning
Preprint · arXiv:2606.22151
032025
Narrative-Centered Emotional Reflection: An Early Prototype for AI-Supported Emotional Self-Reflection
Preprint · arXiv:2504.20342
04
Research Experiences
Graduate Research Assistant
Wayne State University, Trustworthy AI Lab
Advisors: Dongxiao Zhu and Sooin Kim
Research focus 01
AI-Supported Mentoring Systems
- Develop AI-supported mentoring systems aimed at improving retention and educational experiences for veteran engineering students.
- Build LLM-based mentoring components that deliver structured, context-aware guidance and relevant academic resources.
- Contribute to system prototyping, experimental design, data analysis, and evaluation of AI-generated support.
05
Research Projects
Research project 01
Fragile Reasoning: A Mechanistic Analysis of LLM Sensitivity to Meaning-Preserving Perturbations
View preprint- 01
Evaluation
Evaluated Mistral-7B, Llama-3-8B, and Qwen2.5-7B on 677 paired GSM8K problems, revealing answer-flip rates of 28.8%-45.1% under meaning-preserving perturbations.
- 02
Framework
Developed the Mechanistic Perturbation Diagnostics (MPD) framework, integrating logit-lens analysis, activation patching, component ablation, and the novel Cascading Amplification Index (CAI).
- 03
Findings
Identified localized, distributed, and entangled failure modes, demonstrating substantial architectural differences in failure localization and recoverability.
- 04
Repair methods
Evaluated steering vectors and layer fine-tuning as targeted repair methods using Python, PyTorch, Hugging Face Transformers, GPUs, and SLURM.
Research project 02
Novelty-Aware Agentic Retrieval: Comparing Research Contributions Through Structured Multi-Step Reasoning
View preprint- 01
System
Built an agentic retrieval system over a 100-paper corpus using six components: query analysis, iterative retrieval, ranking, contribution extraction, comparison, and answer generation.
- 02
Comparison
Designed structured contribution records and a three-pass comparison agent to identify paper-level overlaps, methodological differences, and research gaps.
- 03
Gap analysis
Developed a problem-method gap matrix that generates citation-grounded evidence for unexplored combinations within a research corpus.
- 04
Results
Achieved a mean Precision@5 of 0.980, nDCG@5 of 0.739, and 84.0% schema compliance across ten evaluation queries.
- 05
Implementation
Implemented the system using Python, GPT-4o, FAISS/Chroma, SentenceTransformers, and structured prompting.
06
Technical Skills
Machine Learning & LLMs
PyTorch, TensorFlow, Hugging Face Transformers, OpenAI API, prompt design, model evaluation, reasoning analysis
Interpretability & Reliability
Logit lens, activation patching, component ablation, steering vectors, perturbation analysis, failure diagnosis
Agentic AI & Retrieval
RAG, FAISS, Chroma, BM25, SentenceTransformers, ReAct-style agents, structured prompting
Programming & Data
Python, Java, JavaScript, TypeScript, SQL, Bash, pandas, NumPy, Matplotlib
Computing & Development
SLURM, HPC, CUDA/GPU computing, Linux, Git, Docker, FastAPI, Flask, React, Vercel
07
Education
Wayne State University
Detroit, Michigan
Ph.D. in Computer Science
Graduate Research Assistant, Trustworthy AI Lab
Boston University
Boston, Massachusetts
Master of Science in Computer Science
Soochow University
Taipei, Taiwan
Bachelor of Arts in Political Science
Contact