I am a fifth-year PhD candidate in the Department of Computer Science at the University of Rochester, advised by Prof. Chenliang Xu. Previously, I spent one wonderful year as a research assistant at the Chinese University of Hong Kong, working with Prof. Chi-Wing Fu on 3D vision. I received my B.Eng. from ESE Department, Nanjing University in 2019. In my undergrad, I worked with Prof. Zhan Ma on image compression.

I am working on multimodal learning and generation. Recently, I am particularly interested in how to leverage the power of large language models (LLMs) to enhance multimodal understanding and generation.

Research opportunities: I am open to collaborating on research projects. Shoot me an email if you are insterested.

Email  /  CV  /  Google Scholar

DRIFT transfers reasoning from DeepSeek-R1 into QwenVL via gradient-space guidance, improving multimodal reasoning without destabilizing alignment or expensive RL.

A benchmark for evaluating cross-modal capabilities and consistency in omni-language models.

How generative models can improve sound separation across diverse categories with visually-guided training.

No fine-tuning, no task-specific data, just latent inversion + text-conditioned denoising to isolate any sound you describe.

We propose a bag of tricks to boost the adversarial transferability of ViT-based attacks.

Introducing MMPerspective, a comprehensive benchmark for MLLMs on perspective understanding.

π-AVAS is a two-stage framework that combines physics-based vision-guided audio simulation for generalization with flow-matching audio refinement for realism.

Where and why you should care about frequency space in diffusion models.

We learn from movies to transform audio to deliver appropriate highlighting effects guided by the accompanying video.

A survey on the recent Large Language Models for video understanding.

We introduce VidComposition, a benchmark designed to assess MLLMs' understanding of video compositions

We use pretrained text-guided diffusion models to scale up/down concepts in image/audio.

A new take on the audio-visual separation problem with the recent generative diffusion models.

We achieve joint audio-visual editing under language guidance.

Thinking of the equivalent of 3D Gaussian Splatting and volumetric primitives for the human body soundfield? Here, we introduce Acoustic Primitives.

We propose a novel method of synthesizing real-world audio-visual scenes at novel positions and directions.

We explore the problem of sound source visual localization in egocentric videos, propose a new localization method and establish a benchmark for evaluation.

A non-local attention based method for point cloud denoising in both synthetic and real scenes.

An image compression system under extreme condition, e.g., < 0.05 bits per pixel (bpp).

• TrustFM: Workshop on Trustworthy Foundation Models, ICCV 2025 — Organizer

• CVPR (2023–2025)
• ICCV (2025)
• AAAI (2023–2025)
• ACM MM (2023–2025)
• NeurIPS (2025)

• IEEE Transactions on Multimedia (TMM)
• IEEE Transactions on Image Processing (TIP)
• SIGGRAPH (ACM)

The template is based on Jon Barron's website.