Bingyu Li · Ph.D. Student at USTC
I am a Ph.D. student at the University of Science and Technology of China (USTC), supervised by Prof. Xuelong Li. My research focuses on open-world visual perception, world action models, and generative video models. The long-term goal behind these directions is simple: Building intelligence for an open and dynamic world.
From recognizing what exists to modeling how the world changes.
Fine-grained visual understanding under open vocabulary and complex conditions.
02 · ModelAction-conditioned representations for dynamics, interaction, and future outcomes.
03 · GenerateEfficient and controllable synthesis for dynamic scene simulation.
Research identity and long-term motivation.
My research explores how machines can perceive open-ended scenes, associate semantic entities through time, model action-conditioned dynamics, and generate plausible futures.
I am particularly interested in connecting these capabilities into a coherent framework for visual intelligence, spanning fine-grained perception, temporal reasoning, world modeling, and controllable generation.
Recent publications and milestones.
Projects organized by the questions they address.
* denotes equal contribution. A complete list is available on Google Scholar.
MTRefSeg introduces Multi-temporal Referring Segmentation (MTRS), where models compare temporally related images and segment the language-described changed region. The work builds MTRefSeg-21K with 21K bi-image-text-mask annotations and proposes MTRefSeg-R1, a change-aware LVLM trained by vision-only temporal pretraining followed by referring multi-temporal fine-tuning.
A benchmark and baseline for realistic open-vocabulary remote sensing segmentation.
Efficient open-vocabulary segmentation methods tailored to remote sensing imagery.
A benchmark and method for open-vocabulary instance segmentation in marine environments.
Training-free segmentation methods for underwater scenes.
A framework for integrating arbitrary visual modalities to improve multimodal semantic segmentation.
An unbiased multiscale modal fusion framework for multimodal semantic segmentation.
Quantitative and spatial awareness for zero-shot object counting in complex scenes, supporting robust perception under open-ended visual conditions.
Fine-grained pixel-text alignment for open-vocabulary semantic segmentation.
Reward-aware trajectory shaping for efficient few-step visual generation across image and video generation settings.
Action-conditioned world modeling for understanding interactions, dynamics, and future outcomes.
Selected academic recognition.
National Scholarship for Graduate Students, 2025.
AAAI 2026 Oral and ACM MM 2025 Oral.
Peer review for journals and conferences.
IEEE TPAMI, IEEE TNNLS, IEEE TGRS, Pattern Recognition, and related journals.
CVPR, NeurIPS, ICLR, and other major conferences.
