01 / FOUNDATIONS
LLM Safety & Alignment
Jailbreak vulnerabilities, robust refusal behaviors, anomalous tokens, and safety evaluation of language models.
RESEARCHER AT THE INTERSECTION OF AI & SECURITY
Hello, I'm
Zihui Wu.
吴子辉 Ph.D. · Postdoctoral Researcher
Exploring how to make intelligent systems safer, more reliable, and better connected — from large language models to networks of autonomous agents.
FIG. 01 / RESEARCH LANDSCAPEAI × NETWORKS
✳Making intelligent systems
trustworthy by design.01—03
trustworthy by design.01—03
02AAAI 2026 papers
01COLING 2025 paper
∞Questions worth asking
FOCUS
Trustworthy, scalable AI
01 / ABOUT ME
Curiosity meets rigor.
I am a Postdoctoral Researcher at Southwest University in Chongqing, China. I received my Ph.D. from Xidian University, where I worked on the security and alignment of large language models.
My research bridges trustworthy AI and machine learning: understanding vulnerabilities in language-model-based systems, and developing scalable graph and tensor learning approaches for complex interactions.
I'm particularly interested in the emerging space where people, AI agents, and networks meet.
02 / RESEARCH INTERESTS
Problems I think about.
From model-level behavior to the structure of interacting intelligent systems.
02 / INTERACTIONS
Trustworthy AI Agents
Security, coordination, and behavioral risk in agentic systems and human–agent interaction networks.
03 / METHODS
Graph & Tensor Learning
Scalable representation learning, high-order interactions, low-rank modeling, and graph-based detection.
✳ Current interests evolve; the publications below document completed work.
03 / SELECTED PUBLICATIONS
Ideas into papers.
01 / 2026
AAAI 2026AI Alignment
HumorReject: Decoupling LLM Safety from Refusal Prefix via a Little Humor
A different approach to safe refusal that uses context-appropriate humor to reduce dependence on explicit refusal prefixes.
AAAI '26
02 / 2026
AAAI 2026Technical Track
GlitchMiner: Mining Glitch Tokens in Large Language Models via Gradient-based Discrete Optimization
Behavior-driven discovery of anomalous LLM tokens through predictive entropy and gradient-guided search.
AAAI '26
03 / 2025
COLING 2025LLM Security
The Dark Side of Function Calling: Pathways to Jailbreaking Large Language Models
An investigation of jailbreak risks introduced by LLM function-calling interfaces, with a focus on safety boundaries.
COLING '25
04 / ACADEMIC JOURNEY
Where I've been.
2026 — PRESENT
Postdoctoral Researcher
Southwest University · Chongqing, China
Trustworthy AI, graph and tensor learningCOMPLETED 2026
Ph.D. in Computer Science
Xidian University · Xi'an, China
Research on LLM security and alignment · Advisor: Prof. Haichang Gao05 / GET IN TOUCH
Let's explore what's next.
Interested in trustworthy AI, emerging agent ecosystems, or scalable machine learning? I'd love to exchange research ideas.