Xumin Yu

I am currently a senior researcher at Tencent Hunyuan, working on large multimodal models and physical AI foundation models.

I obtained my Ph.D. degree from the Intelligent Vision Group (IVG), Department of Automation, Tsinghua University in 2025, advised by Prof. Jiwen Lu and Jie Zhou. Before that, I received my B.Eng. degree from the Department of Electronic Engineering, Tsinghua University in 2020.

I am broadly interested in large language model and computer vision. My current research focuses on multi-modal large language models and large vision models.

Email  /  Google Scholar  /  Github

Publications

* indicates equal contribution

ViQ is a framework for discrete visual representations that balances semantics and details while supporting native-resolution inputs, serving as a unified, general discrete representation for arbitrary visual inputs for both understanding and high-fidelity reconstruction.

HY-Embodied-0.5 is the first version of our embodied foundation models for real-world agents, attaining best performance on 16 out of 22 widely used benchmarks of visual perception, spatial intelligence and embodied reasoning.

P2P is a framework to leverage large-scale pre-trained image models for 3D point cloud analysis.

Point-BERT is a new paradigm for learning Transformers in an unsupervised manner by generalizing the concept of BERT onto 3D point cloud data.

PoinTr is a transformer-based framework that reformulates point cloud completion as a set-to-set translation problem.

We propose a new contrastive regression (CoRe) framework to learn the relative scores by pair-wise comparison, which highlights the differences between videos and guides the models to learn the key hints for assessment.

We propose a graph interaction networks (GINs) model for transferring relation knowledge across two graphs two different scenarios for video analysis, including a new proposed setting for unsupervised skeleton-based action recognition across different datasets, and supervised group activity recognition with multi-modal inputs.

Our method allows machine learning algorithms to perform fine-grained estimation of physiological states (e.g., sleep depth) even if the training labels are coarse-grained.

Website Template