Hi there! 👋 I am Yaxin Luo.
About Me
Decorative animated background
Hello! I am a First-Year Machine Learning PhD student at MBZUAI, advised by Prof. Zhiqiang Shen and Prof.Mohsen Guizani. I am also closely working with my friend Xiaofu Chen. My research aims to develop unified and physical grounded mutlimodal foundation models and make them efficient and faster, be deployable on edge devices — and, more ambitiously, to enable training them directly using edge devices’ computing power. I wish to approach this challenge from a co-design perspective across the data, algorithm, system, and hardware.
Previously, I earned my Bachelor’s degree from Technical University of Denmark, where I was fortunate to be supervised by Prof. Dim P. Papadopoulos. Meanwhile, I was lucky to collarating with Dr.Gen Luo and Prof.Rongrong Ji on efficient deep learning researches during my bachelor. Earlier, I spent an intense and rewarding year at the University of Edinburgh studying pure mathematics and physics—an experience that sparked my passion for science and technology, deepened my curiosity about the unknown, I was curious and wanted to explore String Theory at that time, this one year ultimately shaped who I am today. Before Edinburgh, while enrolled in a Bio-Medicine program at the University of Queensland and preparing for the UCAT test to be addimitted into the university's medical school, I failed at the end. As I only focused on managing a high-street multi-brand boutique which was located in Brisbane‘s Southbank near the casino, and was far more focused on business than on study and research; that Edinburgh year changed my priorities and set me on a research path, thanks to the advice, encourage and supports of my academic personal tutor Prof.Ana Rita Pires when I was at Edinburgh. Anyway, all those past experiences have made me who I am today.
My research interests focus on:
- Efficient Machine Learning : Beyond the perspectives of large scale, I enjoy compressing the large foundation models into real-world application deployable ones, from both data , algorithm–system and hardware full stack co-design perspectives. as well as considering all training and inference stages techiniques. My aim is to deploy the advanced AI models to edge devices that we used and faced in the daily life, such as, phones, laptops and drones etc..
- Multimodal Foundation Model : Developing native multimodal foundation models which can perform unified understanding, reasoning, generation tasks from video, language, speech. These models will serve as the core intelligence—the "brain"—for Embodied AI, Robotics, and many other applications.
- Physics Grounded Foundation Model : Vision-centric video model that learns causal structure and explicit physics knowledge from large-scale videos supports both understanding and generation for physical real-world; and further enabling action-conditioned prediction for embodied agents.
Recently, I am focusing on LLM's training data anatomy and on-device unified understanding & generation model.
News
[2025-09-18] 🚀 OpenCaptchaWorld has been accepted by NeurIPS 2025.
Selected Publications
( * indicate equal contribution)
For full and up-to-date publication list, please refer to my Google Scholar page.
OpenCaptchaWorld: A Comprehensive Web-based Platform for Testing and Benchmarking Multimodal LLM Agents
APL: Anchor-Based Prompt Learning for One-Stage Weakly Supervised Referring Expression Comprehension
γ-MoD: Exploring Mixture-of-Depth Adaptation for Multimodal Large Language Models
DViN: Dynamic Visual Routing Network for Weakly Supervised Referring Expression Comprehension