✨ About me

Bonjour, I am a Ph.D. student in the Department of Computer Sciences at the University of Wisconsin–Madison, advised by Prof. Sharon Li. Before this, I obtain the graduate degree of computer application technology at Peking University, advised by Prof. Jie Chen. I received a B.S. degree from the College of Computer Science (Elite Class), in Jilin University, in 2022.

My research focuses on Trustworthy Multimodal Intelligence, with particular interests in:

Interpretability and Mechanistic Alignment: Understanding the internal representations and decision-making mechanisms of multimodal models, with the goal of making their reasoning more interpretable, controllable, and safety-aligned.
Agentic Image and Video Generation: Building capable and efficient generative agentic systems through post-training (e.g., SFT, preference optimization, and reinforcement learning) and agent harnesses that enable planning, tool use, iterative feedback, and self-correction for high-fidelity and efficient visual generation.
Unified and Omni Multimodal Models: Developing unified multimodal models that integrate perception, reasoning, generation, and action within a shared framework, toward more general-purpose multimodal intelligence.

😃 I am seeking research collaboration opportunities. Feel free to contact me if you are interested in potential collaboration!

📰 News

2026/04: 🎉 One paper about Multimodal Interpretability is accepted by ACL 2026 as main paper.
2024/12: 🎉 One paper (AnyTalk about Talking Head Generation is accepted by AAAI 2025.
2024/09: 🎉 One paper (SubgDiff about molecular 3D conformation generation is accepted by NeurIPS 2024.
2024/07: 🎉 Two papers (Adashield about VLM Safety and ParCo about Text-to-Motion Synthesis) are accepted by ECCV 2024.
2024/04: 🚀 We release an open-source repository about Awesome-T2I-safety-Papers.
2023/10: 🎉 One paper (TIDA) about open-set semi-supervised learning is accepted by NeurIPS 2023.
2023/03: 🎉 Two papers (OSP about robust semi-supervised learning and FPR about semi-supervised semantic segmentation) are accepted by CVPR 2023.

💻 Selected Publications

My complete publication record is available on Google Scholar. Each highlighted paper below includes a visual overview and direct links to its resources.

Cross-branch semantic steering pipeline transferring semantic vectors from multimodal understanding to image generation.

arXiv 2026Multimodal Interpretability

Do Unified Multimodal Models Think in One Space? A Lens Through Cross-Branch Steering

Yu Wang, Sharon Li

An intervention-based study of whether multimodal understanding and generation share a transferable semantic space.

Paper

Visual overview of multimodal in-context learning setups, performance comparison, and task-mapping transfer.

ACL 2026Multimodal Interpretability

Why Multimodal In-Context Learning Lags Behind? Unveiling the Inner Mechanisms and Bottlenecks

Yu Wang, Sharon Li

A mechanistic study of how multimodal in-context learning routes visual and textual evidence—and where that reasoning pipeline falls short.

Paper coming soon · Code coming soon

AdaShield defense pipeline against structure-based multimodal jailbreaks.

ECCV 2024Safety AI

AdaShield: Safeguarding Multimodal LLMs from Structure-based Attack via Adaptive Shield Prompting

Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, Chaowei Xiao

Adaptive prompt shielding for defending multimodal language models against structure-based jailbreak attacks.

Paper · Code

Contextual taxonomy priors for open-world semi-supervised learning.

NeurIPS 2023Robust Semi-supervised Learning

Discover and Align Taxonomic Context Priors for Open-world Semi-Supervised Learning

Yu Wang, Zhun Zhong, Pengchong Qiao, Xiawu Zheng, Chang Liu, Nicu Sebe, Rongrong Ji, Jie Chen

Learning taxonomic context priors to discover and align known and novel classes in open-world semi-supervised learning.

Paper · Code

Out-of-distribution semantic pruning for robust semi-supervised learning.

CVPR 2023Robust Semi-supervised Learning

Out-of-Distributed Semantic Pruning for Robust Semi-Supervised Learning

Yu Wang*, Pengchong Qiao*, Chang Liu, Guoli Song, Xiawu Zheng, Jie Chen

A semantic-pruning approach that improves the reliability of semi-supervised learning under distribution shift.

Paper · Code

AnyTalk multi-modal and multi-domain talking head generation framework.

AAAI 2025AIGC

AnyTalk: Multi-modal Driven Multi-domain Talking Head Generation

Yu Wang, Yunfei Liu, Fa-Ting Hong, Meng Cao, Lijian Lin, Yu Li

A multimodal framework for controllable talking-head generation across diverse identities and domains.

Paper

🎖 Honors and Awards

2025/05: Outstanding Graduate, Peking University (University Graduation Honor) (Top 5%)
2024/11: Tiehan Scholarship, Peking University (Top 5%)
2024/11: Merit Student, Peking University (Top 10%)
2023/11: Merit Student, Peking University (Top 10%)
2023/03: Excellent Papers of China Computer Federation Computer Application Conference, CCF Conference on Computer Applications (Top 1%)
2022/12: Jilin Bank Wang Xianghao Scholarship, Jilin University (Top 1%)
2021/11: National Scholarship, Jilin University (Top 1%)
2020/10: National Inspirational Scholarship, Jilin University (Top 1%)

💪 Academic Service

PC Member: CVPR’24-26, ICLR’24-26, ICML’24-26, ACM MM’24, ECCV’24,26, NeurIPS’24-25, AAAI’26

If you like the template of this homepage, welcome to star and fork the open-sourced template version AcadHomepage .