research

Vision-language model robustness

Physical Superintelligence Lab, University of Southern California
Research Assistant · June 2025–Present

I investigate how vision-language models fail and how targeted post-training can improve their robustness and reliability. My work includes:

• Developing pipelines to generate, filter, and evaluate multimodal data across visual reasoning tasks.
• Analyzing failures in object recognition, attributes, spatial relationships, counting, and negation.
• Applying large vision-language models and parameter-efficient fine-tuning to study targeted repair and generalization.

Related work is listed on my publications page, including Large Reward Models and FuzzingRL.

Microfabrication and quantum transport

Laboratory of Microfabrication, Institute of Physics, Chinese Academy of Sciences
Research Assistant · May–August 2023

I assisted with microelectronic substrate fabrication for terahertz communication devices using lift-off processes. I operated mask aligners, ICPCVD systems, and electron-beam evaporation tools, and supported ultra-low-temperature experiments and quantum transport measurements.