I study how large language and multimodal models behave under challenging reasoning, safety, and adaptation settings, with a focus on robustness, red-teaming, and in-context learning.
Adversarial evaluation, jailbreaks, red-teaming.
Finding and understanding failure modes in increasingly capable models.
Tool use, multi-agent decisions, delayed feedback.
Multimodal adaptation from visual context.
Education
Ph.D. Candidate, Computer Science
M.Sc. Data Science
Experiences
Doctoral Researcher
Applied Scientist Intern
Recent News
View all- We won the Autoresearch Challenge at Paris Research Hackathon 2026.
- One co-first paper on red-teaming implicit vulnerabilities of T2I models was accepted to ECCV 2026.
- I received a Silver Reviewer Award from ICML 2026.
- I received a compute grant from CAIS to support research on AI safety.
- Deep Research Brings Deeper Harm is online.
Publications
View allTrue Multimodal In-Context Learning Needs Attention to the Visual Context
Shuo Chen, Jianzhe Liu, Zhen Han, Yan Xia, Daniel Cremers, Philip Torr, Volker Tresp, Jindong Gu.
Bag of Tricks for Subverting Reasoning-based Safety Guardrails
Shuo Chen, Zhen Han, Haokun Chen, Bailan He, Shengyun Si, Jingpei Wu, Philip Torr, Volker Tresp, Jindong Gu.
Contact
I am looking for motivated Bachelor’s and Master’s students who are willing to publish high-impact papers on top venues for research projects on LLMs, MLLMs, and Agents. Previous supervisions have led to work at ECCV, EMNLP, COLM, and ACL.
A=LastNameFirstName dot cs B=outlook; A@B.com