Wenzhang DuAI safety & empirical machine learning
中文Let’s talk

INDEPENDENT RESEARCHER

Wenzhang Du.

AI moves fast.
I work to close the gaps.

As AI accelerates research, rigorous validation needs to keep pace. I work to uncover blind spots, test assumptions, and address hidden failures through model auditing, statistical evaluation, and empirical research.

Explore my researchGet in touch

Daqing, ChinaB.Eng. in Computer Engineering and Artificial Intelligence

Scroll to discover
Publications & accepted papers
15 / 2024–2026
Research directions
03 / AI · ML · EO
Best Paper · ICPEI
2025
Google ScholarGitHub

01 / ABOUT

Better questions.
More reliable answers.

My research starts with a simple question: how do we know an AI system is behaving reliably?

I am an independent researcher and a B.Eng. graduate of the International College at Mahanakorn University of Technology. My interests center on LLM safety, trustworthy AI, AI agent evaluation, and rigorous empirical machine learning.

I am interested in practical ways to uncover failure cases, test prompt and instruction robustness, detect hallucinations and risk, and compare models with statistical care. My background also spans efficient deep learning, computer vision, and remote sensing.

For graduate research, I hope to build evaluation frameworks that make AI system behavior easier to test, compare, and improve.

GitHub Google Scholar

02 / RESEARCH

Three connected
research directions.

From understanding failure to measuring progress and putting reliable learning into practice.

01

Trustworthy AI

Understanding how language models and AI agents fail, and how to make their behavior easier to audit.

  • LLM safety
  • Agent evaluation
  • Model auditing
02

Rigorous evaluation

Building careful comparisons through paired tests, bootstrap confidence intervals, and reproducible experiments.

  • Robust empirical ML
  • Failure analysis
  • Service-level trust
03

Learning in the real world

Exploring efficient models and reliable learning for computer vision, Earth observation, and resource-constrained devices.

  • Remote sensing
  • Computer vision
  • Efficient deep learning

SELECTED WORK

Full publication index
01
WSDM 2027 ·

Complex Utility, Simple Calibration: Decision Sufficiency in RAG Evidence Selection

02
AACL-IJCNLP 2026 ·

PAIRSAFE: Proxy-Resistant Paired Safety Evaluation for Refusal Boundaries

03
PAKDD 2026 ·

When +1% Is Not Enough: A Paired Bootstrap Protocol for Evaluating Small Improvements

04
WCCI (CEC) 2026 ·

ARC-CTRL: Audit-Guided Closed-Loop Trust Control for Surrogate-Assisted Evolutionary Hyperparameter Optimization

03 / PUBLICATIONS

The publication index.

Conference publications and accepted papers, spanning trustworthy learning, evaluation, and applied AI.

Google Scholar

Showing 15 of 15 publications

2026 — 2024

Major publications

7 papers

2026

7

EI conference papers

8 papers

2026

3

2025

4

2024

1

04 / EXPERIENCE

A foundation built
through research.

Education

— Graduated

B.Eng. in Computer Engineering and Artificial Intelligence

International College · Mahanakorn University of Technology

Cumulative GPA3.5 / 4

05 / RECOGNITION

Recognition
& support.

Best Paper Award

ICPEI 2025

Project Day Recognition

Mahanakorn University of Technology · Project Day

Research travel support

IEEE CIS Travel Grant · PAKDD Travel Grant · IGARSS Travel Grant

Tools & methods

Programming & research
Python, C++, Java, experiment scripting, data processing, reproducible workflows, LaTeX
Safety & auditing
Deep learning, trustworthy AI, LLM safety evaluation, AI agent auditing, hallucination and risk analysis
Statistical evaluation
Paired tests, bootstrap confidence intervals, multi-seed experiments, metric design, uncertainty reporting
Applied machine learning
Computer vision, remote sensing, model compression, quantization-aware learning, edge-aware ML

06 / CONTACT

Good research starts
with a conversation.

Interested in trustworthy AI, evaluation, or graduate research? I would be glad to exchange ideas.

dqswordman@gmail.com
GitHubGoogle Scholar

Daqing, China

Copy text

Copy was unavailable. Select and copy the text below.