Xueyi Ke (柯学奕)
Hello! I'm Xueyi, a PhD student in Computer Science at UC Irvine.
My research focuses on how foundation models represent and compute internally, and how this understanding can help build more efficient, interpretable, and self-improving systems. My previous work spans representation analysis in multimodal and vision-language-action models, multimodal learning, and LLM post-training.
I earned my undergraduate and master's degrees from Wuhan University and Nanyang Technological University. I have also interned at Tsinghua University and Alibaba Group.
Research
-
Reasoning & Self-Improvement
Latent reasoning, self-distillation, post-training, and self-improving agents.
-
Representations & Interpretability
Understanding what foundation models encode, and how internal computation gives rise to model behavior.
-
Multimodal & Agentic Systems
Vision-language models, VLA, multimodal learning, and sequential decision systems.
Selected Research
Projects
-
VLA Mechanistic Interpretability
2025 – 2026Tracing how vision-language-action models turn visual observations into action-token predictions — attention circuits, neuron-level mechanisms of embodied decision-making, and the recovery of induction-head behaviour in one-shot imitation learning.
-
PassGPT — Personalized Education Agent
2024An educational QA agent built on retrieval over roughly 50K curated documents, with SFT and RLHF applied to about 100K domain-specific samples.
Experience
Nanyang Technological University
Research Associate
Aug 2025 - May 2026
Singapore
Alibaba Group
Algorithm Research Intern
Apr 2025 - Aug 2025
Beijing, China
Tsinghua University
Research Intern
May 2023 - Aug 2023
Beijing, China
Education
University of California, Irvine
Ph.D. in Computer Science
Sept 2026 - Present
Irvine, CA, USA
Nanyang Technological University
M.Sc. in Computer Control & Automation
Aug 2023 - Aug 2025
Singapore
Wuhan University
B.Eng. in Electrical Engineering & Automation
Sept 2019 - June 2023
Wuhan, China