Yongxue Xu (Jerry)

I am an undergraduate student at Sun Yat-sen University, majoring in Intelligent Science and Technology in the School of Intelligent Systems Engineering. My current research interests lie in video generation, world models, 4D scene understanding, and world-action models (WAM).

My research journey began in a data structures course, where I was first drawn to the elegance of representing a problem through structure. This taste later led me to graph-based decision problems. In robust competitive influence maximization on multi-layer graph networks and critical node identification in urban road networks, I was interested in how structure shapes behavior: how information propagates across layers, how a few key nodes can change a system, and how a model should make decisions when the topology itself carries most of the meaning. These projects made me care less about isolated predictions and more about the hidden state of a system - the relationships, constraints, and dynamics that make an output meaningful.

Later, when I moved into video instance segmentation and 4D tracking, the same concern reappeared in a more visual form. A mask, trajectory, or generated frame can be plausible at one moment and still fail once the camera moves, an object is occluded, or the scene is revisited. I gradually realized that long-video intelligence is not only a recognition problem; it is a memory problem. Instead of asking only whether a model can render the next frame, I became more interested in whether it can remember that it is still rendering the same world.

This realization led me toward video generation and action world models. I believe the missing piece is not simply a sharper decoder, a longer context window, or a larger generative backbone, but a persistent object-state core: a compact, structured, updatable memory of which entities exist, where they are, how they move, how they relate, and how they should persist through time, occlusion, viewpoint change, and action. A world model that cannot remember is not really modeling a world; it is a renderer with momentum.

More broadly, I want to build models in which perception, prediction, generation, and control become different operations on the same state. Understanding should update the core from new observations; prediction should roll the core forward; generation should render the core into pixels; control should edit the core and make the change stay; and action should write its consequences back into the world. I am currently working closely with collaborators at the EPIC Lab of SJTU SAI, HKUST, ByteDance, Xiangru Huang's Embodied AI Lab at Westlake University, and PKU LumPin Lab on video generation, WAM, and multimodal spatiotemporal understanding projects, where these questions become concrete systems rather than only conceptual interests. My current research is centered on controllable video generation and interactive world models.