Hanchen Xie (谢瀚辰)

I am a Research Scientist at Huawei 2012 Labs, where I work on agent frameworks. My research focuses on how agents construct and manage context and memory, how multimodal information can be integrated effectively, and how systems can learn from their own operation.

I received my Ph.D. in Computer Science from the University of Southern California, where I was based at the Information Sciences Institute. My earlier research focused on learning with limited supervision — including semi-supervised and few-shot image classification — and on learning representations that recover disentangled and causal structure without dense supervision.

The later part of my Ph.D. research grew from a question I explored in my ICML 2023 paper: Do vision-based models that appear to learn object dynamics remain reliable when the environment changes? Often they do not. I framed environment misalignment as a problem distinct from conventional domain shift, introduced datasets designed to isolate it, and quantified its impact on state-of-the-art models.

My subsequent work traced these failures to specific architectural choices and introduced symbolic alignment: a class-agnostic symbolic space shared across domains, allowing relations learned in one environment to transfer to another. More recently, I extended this perspective to large vision models and video. Together, this body of work formed the foundation of my dissertation.

Research interests

• Agent frameworks — how an agent's context and memory should be constructed and managed, what place multimodal information should hold within them, how tools and the execution harness are exposed, and how a system improves from its own operation.
• Environment misalignment and symbolic alignment — why models degrade once the visual environment changes, and how a shared symbolic space preserves the relations they learned.
• Representation learning — recovering disentangled and causal structure from weak supervision rather than dense labels.
• Learning with less labels — semi-supervised and few-shot learning when annotation is the scarce resource.

Publications

• 
  Efficiently Mitigating Video Content Misalignment on Large Vision Model with Time-Series Data Alignment

  Hanchen Xie, Rui Ma, Jiageng Zhu, Zheng Mai, Wael Abd-Almageed, Zubin Abraham

  CVPR 2025 Workshop on Efficient Large Vision Models (eLVM)
• 
  Look, Learn and Leverage (L³): Mitigating Visual-Domain Shift and Discovering Intrinsic Relations via Symbolic Alignment

  Hanchen Xie*, Jiageng Zhu*, Mahyar Khayatkhoei, Jiazhi Li, Wael AbdAlmageed

  Preprint, 2024
• 
  An Investigation on The Position Encoding in Vision-Based Dynamics Prediction

  Jiageng Zhu*, Hanchen Xie*, Jiazhi Li, Mahyar Khayatkhoei, Wael AbdAlmageed

  European Conference on Computer Vision (ECCV) Workshops, 2024
• 
  A Critical View of Vision-Based Long-Term Dynamics Prediction Under Environment Misalignment

  Hanchen Xie, Jiageng Zhu, Mahyar Khayatkhoei, Jiazhi Li, Mohamed E. Hussein, Wael AbdAlmageed

  International Conference on Machine Learning (ICML), 2023
• 
  Partner-Assisted Learning for Few-Shot Image Classification

  Jiawei Ma*, Hanchen Xie*, Guangxing Han, Shih-Fu Chang, Aram Galstyan, Wael Abd-Almageed

  International Conference on Computer Vision (ICCV), 2021
• 
  MUSCLE: Strengthening Semi-Supervised Learning via Concurrent Unsupervised Learning Using Mutual Information Maximization

  Hanchen Xie, Mohamed E. Hussein, Aram Galstyan, Wael Abd-Almageed

  Winter Conference on Applications of Computer Vision (WACV), 2021

Education

• Ph.D., Computer Science — University of Southern California2025
• M.S., Computer Science — University of Southern California2019
• B.S., Electrical Engineering — Michigan State University2017

Academic service

Misc

Outside of research, I ski, play tennis, and spend as much of the winter as I can at Mammoth Mountain. I also enjoy road trips and the outdoors, and spent several years riding motorcycles.