I am an undergraduate student at Sun Yat-sen University, majoring in Intelligent Science and Technology. My research focuses on video generation, world models, and multimodal spatiotemporal understanding. My research began with graph learning and has grown toward video and 4D scenes. I am particularly interested in how models represent objects, maintain their states over time, and use generation to understand and interact with the world. I collaborate with researchers at HKUST, Shanghai Jiao Tong University, and Westlake University on video generation and embodied intelligence. I am also part of InkMind.AI, where we explore multimodal understanding and generation, along with tools and workflows for AI-assisted research. I welcome research collaborations in generative modeling and embodied intelligence. Feel free to contact me.
* Equal contribution; † Corresponding author
Submitted to ICLR 2027
@misc{xu2026last,
title={LaST: Latent State Tracking for Persistent Object Memory in Video World Models},
author={Xu, Yongxue and Yang, Harry},
year={2026},
note={Submitted to ICLR 2027. Corresponding author: Harry Yang}
}Submitted to ICLR 2027
@article{kang2026trace,
title={TRACE: Typed Reference-Aware Conversational Editing for Humanoid Motion Diffusion},
author={Kang, Zejian and Luo, Yifu and Shen, Boyang and Li, Jingxuan and Chen, Zekai and Shi, Zekun and Xu, Yongxue and Hou, Zhenyu and Wu, Bingqian and Zhang, Shaolei and Yang, Wentao and Huang, Xiangru},
journal={arXiv preprint arXiv:2610.21515},
url={https://arxiv.org/abs/2610.21515},
year={2026},
note={Submitted to ICLR 2027. Equal contribution: Zejian Kang, Yifu Luo, and Boyang Shen}
}Submitted to ICLR 2027
@article{liu2026evigda,
title={Graph Domain Adaptation Does Not End with Representation Learning},
author={Liu, Ziqian and Xu, Yongxue and Zhang, Enze and Zhang, Jiaqi and Wang, Hao and Wang, Maolin},
journal={arXiv preprint arXiv:2609.25692},
url={https://arxiv.org/abs/2609.25692},
note={Submitted to ICLR 2027. Equal contribution: Ziqian Liu and Yongxue Xu. Corresponding author: Maolin Wang},
year={2026}
}Submitted to AAAI 2027
@article{xu2026posreasoner,
title={Learning to Reason with Persistent Object States for Video Instance Segmentation},
author={Xu, Yongxue and Yang, Boxue and Liu, Ziqian and Zhang, Shaoqiu and Qian, Rui and Chen, Haopeng},
journal={arXiv preprint arXiv:2609.35539},
url={https://arxiv.org/abs/2609.35539},
note={Submitted to AAAI 2027. Equal contribution: Yongxue Xu and Boxue Yang. Corresponding author: Haopeng Chen},
year={2026}
}IEEE Transactions on Network Science and Engineering · Under review
@article{xu2026influence,
title={Solving the Robust Influence Maximization Problem in Competitive Multilayer Networks via a Diffusion-Aware Role-Guided Evolutionary Approach},
author={Xu, Yongxue and Liu, Ziqian and Huang, Yongqing and Wang, Shuai},
journal={Under review at IEEE Transactions on Network Science and Engineering},
note={Corresponding author: Shuai Wang},
year={2026}
}Submitted to NeurIPS 2026
@article{li2026lmmtrack4d,
title={LMM-Track4D: Eliciting 4D Dynamic Reasoning in LMMs via Trajectory-Grounded Dialogue},
author={Li, Chaoyue and Xu, Yongxue and Feng, Jie and Ding, Jiayu},
journal={arXiv preprint arXiv:2605.19390},
note={Equal contribution: Chaoyue Li, Yongxue Xu, and Jie Feng. Corresponding author: Jiayu Ding},
year={2026}
}GBCESC 2025 · Best Paper Award
@inproceedings{liu2025criticalnodes,
title={Identifying Critical Nodes with Deep Learning and Reinforcement Learning: A Case Study on Urban Road Networks},
author={Liu, Zhonghan and Liu, Wenshuo and Xu, Yongxue and Liu, Ziqian and Liu, Yichen and Wang, Shuai},
booktitle={GBCESC},
note={Best Paper Award. Corresponding author: Shuai Wang},
year={2025}
}B.Eng. in Intelligent Science and Technology (in progress)
School of Intelligent Systems Engineering
Working on building Omni-Simple-World, an embodied foundation model that unifies task understanding, world prediction, and action generation. In parallel, pursuing frontier research in Physical Intelligence, exploring embodied reasoning and action generation in generative models.
Research on embodied world-action models (WAMs), with a focus on learning from failures and long-horizon robotic tasks.