Visual Agents and Multimodal Foundation Models
Supervised by Dr. Mingfei Gao and Dr. Mingze Xu
I obtained my Ph.D. degree from the College of Computer Science and Artificial Intelligence at Fudan University (Shanghai, China) in 2026, advised by Dr. Zuxuan Wu at the Fudan Vision & Learning Lab. Before that, I received my B.Sc. in Computer Science from Fudan University in 2021.
During my Ph.D., I interned at Apple, where I worked with Dr. Mingfei Gao and Dr. Mingze Xu. I also interned at Microsoft Research Asia under the supervision of Dr. Qi Dai.
My research focuses on multimodal AI, particularly visual agents and multimodal generation.
* denotes equal contribution.
Preprint, 2026.
We present MintAct, a family of vision-language models that unifies UI grounding, multi-step navigation, and visual tool use across mobile, desktop, and web environments. Scalable environments and asynchronous reinforcement learning enable generalist agents to match domain specialists, achieving a score of 48.9 on OSWorld-Verified.
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.
We present UniGen-1.5 for unified image understanding, generation, and editing. A shared-reward reinforcement learning strategy jointly improves generation and editing, while a lightweight Edit Instruction Alignment stage improves instruction comprehension.
Advances in Neural Information Processing Systems (NeurIPS), 2025.
We introduce UniGen, a unified multimodal large language model for understanding and generation. Our Chain-of-Thought Verification (CoT-V) strategy allows the model to act as both an image generator and a step-by-step image quality verifier at test time.
IEEE/CVF International Conference on Computer Vision (ICCV), 2025.
We present Reducio, an end-to-end framework for efficient, high-resolution video generation. Its VAE uses a spatiotemporal downsampling ratio of 32 × 32 × 4 to compress motion latents.
Advances in Neural Information Processing Systems (NeurIPS), 2024.
We introduce DeepStack, a simple strategy for connecting vision and language transformers. By injecting visual tokens into multiple LLM layers, it improves multimodal understanding.
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023.
We introduce ResFormer to improve vision transformer performance across image resolutions, including unseen resolutions. The approach combines multi-resolution training, a scale consistency loss, and adaptive global-local positional embeddings.
IEEE Transactions on Multimedia (TMM), 2025.
We study how vision transformer design and training affect robustness to common corruptions, identifying architectural choices that improve robustness. We also introduce input-conditioned dynamic data augmentation.
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2026.
We introduce TinyMem for long-form video action detection. It combines succinct textual descriptions with region-condensed visual memory to capture long-range dependencies efficiently.
Visual Agents and Multimodal Foundation Models
Supervised by Dr. Mingfei Gao and Dr. Mingze Xu
Generative AI · Scalable Vision
Supervised by Dr. Qi Dai
Ph.D. in Computer Science and Artificial Intelligence
B.Sc. in Computer Science (Elite Program)
Conference Reviewer: CVPR, ICCV, ICLR, ICML, NeurIPS, and AAAI.
Journal Reviewer: TMLR and TMM.