Hello! I am Yujia Cai, a third-year undergraduate student majoring in Computer Science at Xidian University. My research experience spans Computer Vision and Multimodal Learning, with prior work on retrieval, sketch-based representation, domain adaptation, and medical image analysis.
I am particularly interested in learning robust and efficient representations that generalize across domains and modalities, and in building practical models for vision and multimodal understanding.
🔥 News
- 2026.01: 🎉 Our paper Hystar: Hypernetwork-Driven Style-Adaptive Retrieval via Dynamic SVD Modulation was accepted to ICLR 2026.
📝 Publications
ICLR 2026
Hystar: Hypernetwork-Driven Style-Adaptive Retrieval via Dynamic SVD Modulation
Yujia Cai, Boxuan Li, Chenghao Xu, Jiexi Yan
International Conference on Learning Representations (ICLR), 2026
- We propose Hystar, a hypernetwork-driven style-adaptive retrieval framework that dynamically modulates the singular values of attention layers via hypernetworks to improve robustness under diverse visual styles. The method introduces an OT-weighted StyleNCE loss to reweight hard negatives and mitigate cross-style semantic confusion. Experiments on DomainNet and DSR demonstrate consistent improvements in cross-style retrieval and zero-shot classification.
📖 Educations
- 2023.09 - 2027.06 (now), Bachelor of Science in Computer Science, Xidian University, Xi’an, China
💻 Internships
- Oct 2024 – Present, Research Intern, OPTIC Lab, Xidian University, China.
- Jan 2026 – Present, Research Intern (Remote), Stony Brook University, USA.