Qingyu Meng [tɕʰiŋ'y, mɤŋ]

PhD Researcher
Vrije Universiteit Amsterdam, Amsterdam, The Netherlands

Email: qingyumeng1128 at gmail.com

Hello world! I am Qingyu Meng, and people who know me also call me Beren. Not the Dutch word for "bear", but the protagonist from Tolkien's book (Beren and Lúthien). Feel free to explore my site to learn more about my interests, experience, and skills.

About me

• I am a first-year PhD researcher at Vrije Universiteit Amsterdam, where I work on trustworthy AI, AI security, and mechanistic interpretability. My research aims to apply/develop mechanistic interpretability methods to make foundation model architectures and agentic systems safer and more trustworthy. I'm particularly driven by the principle of actionable interpretability, so moving from understanding AI black-boxes to actively editing and steering their behavior for real-world reliability.
• I received my M.Sc. in Statistics and Data Science from Leiden University, focusing on statistics, data science, and machine learning.
• I completed my B.A. in Celtic Linguistics and Applied Data Science at Utrecht University.
• Before starting my PhD, I worked as a Data Scientist Intern at Rabobank and an IT Architecture Intern at Kraft Heinz.

Research Interests

• Trustworthy Agentic AI
• AI Security and Safety
• Representation Learning
• Mechanistic Interpretability
• Foundation Model Architectures: MoE, dLLM, KAN, Mamba, and beyond
• Neuro-Symbolic Intersections
• Differential Privacy

Selected Publications

• CCS 2026 (CORE A* / CCF A)SEAL: Reinforcing Global Safety in Mixture-of-Experts through Shared Expert ALignment

  Qingyu Meng, Yiwei Zha, Jiahuan Pei, Koen Hindriks, Herbert Bos, Min Chen

  To appear in ACM CCS 2026[PDF] [arXiv] [Website]

• ACL 2026 (CORE A* / CCF A)StoryMI: Steerable Multi-Agent Therapeutic Dialogue Generation

  Qingyu Meng, Min Chen, Dingming Liu, Yifan Mo, Yue Su, Xin Sun, Koen Hindriks, Jiahuan Pei

  In Findings of ACL 2026[PDF] [BibTeX]

• ACL 2026 (CORE A* / CCF A)SciText2Eq: Assessing LLMs for Explainable Equation Generation for Scientific Creativity

  Yifan Mo, Xiao Fu, Yue Su, Qingyu Meng, Koen Hindriks, Qingzhi Liu, Jiahuan Pei

  In Findings of ACL 2026[PDF] [BibTeX]

• INLG 2026 (CORE B)Trust Stack for Mental Health AI: A Survey of Calibration across Human, Interaction, and AI Layers

  Xin Sun, Yue Su, Yifan Mo, Qingyu Meng, Yuxuan Li, Min Chen, Mengyuan Zhang, Saku Sugawara, Charlotte Gerritsen, Sander L. Koole, Koen Hindriks, Jiahuan Pei

  To appear in INLG 2026[PDF] [BibTeX]

See all in my publications.

Education

• Oct 2025 – PhD, Trustworthy Agentic AI
  Vrije Universiteit Amsterdam
  Trustworthy AI, AI security, mechanistic interpretability
• Sep 2023 – Jul 2025 MSc, Statistics and Data Science
  Leiden University
  Statistics, data science, machine learning
• Sep 2020 – Jul 2023 BA, Celtic Linguistics and Applied Data Science
  Utrecht University
  Linguistics, applied data science

Experience

• Oct 2025 – PhD Researcher
  Vrije Universiteit Amsterdam
  Trustworthy AI, AI security, mechanistic interpretability
• Mar 2025 – Jul 2025 Data Scientist Intern
  Rabobank, The Netherlands
  LLM-based transaction categorization, PySpark pipelines, synthetic data methods, ML model development
• Mar 2024 – Oct 2024 IT Architecture Intern
  Kraft Heinz, Amsterdam
  Data modeling (C4 model), ServiceNow dashboards, architecture runway

Skills & Languages

Auch uns Wanderer führt jeder Weg nach Hause.

To learn something about anything, to learn anything about something.

Far better an approximate answer to the right question, which is often vague, than the exact answer to the wrong question, which can always be made precise.

The Master said, "The 300 verses of the Book of Odes can be summed up in a single phrase: 'Without depraved thoughts.'"

Die Grenzen meiner Sprache bedeuten die Grenzen meiner Welt.