Teaching agentsto think.

My research interests center on post-training for large models, multimodal agents, and physics-constrained spatiotemporal forecasting and understanding. I focus on reinforcement-learning algorithms for large models, coding agents, efficient visual representation learning, and multimodal intent understanding.

Beyond academic publications, I actively participate in research projects and production-grade system development. I contribute to “Regional High-Resolution Intelligent Ensemble Forecasting with Probabilistic Diffusion Models” (National Natural Science Foundation of China, No. 4250050812). To address the limited ability of data-driven methods to quantify atmospheric uncertainty in regional high-resolution ensemble forecasting, the project develops a physics-constrained probabilistic diffusion framework. It incorporates dynamical conservation laws into prior-distribution modeling for meteorological variables and uses global large-scale forecasts to condition the sampling process, enabling the efficient generation of reliable initial fields for ensemble prediction.

I also contribute to “Multilevel Anomalous-Event Detection for Unlabeled Video Data” (National Natural Science Foundation of China, No. 6250075854). To address the limited ability of existing unlabeled-video anomaly-detection methods to capture multilevel feature differences associated with anomalous events, the project develops a multilevel feature-distillation approach that integrates pretrained foundation models. It distills knowledge of normal events from unlabeled videos to construct a normality model and identifies anomalous events through multilevel feature discrepancies.

In university–industry system development, I have contributed to a robot-assisted guidance system for skull-base foramen ovale puncture. As a member of a medical-assistance collaboration with the provincial science and technology department, I optimized methods for localizing the foramen ovale target and skin entry point; the system passed acceptance review by the Second Bethune Hospital of Jilin University. Designed to support robot-assisted navigation for trigeminal nerve puncture procedures, the system completes 3D reconstruction in 15 seconds, treatment-planning simulation in 30 seconds, and the treatment procedure in 20 minutes, substantially reducing treatment time. Its noninvasive calibration avoids secondary injury, supports rapid recovery, and minimizes disruption to patients’ daily lives and work. I also contributed to the Jilin Provincial Cultural Tourism Aggregation Platform. As a core full-stack developer, I implemented full-stack logic for government data review and the aggregation of monthly and annual statistical reports. The platform currently serves more than 300 tourist attractions and over 840 related businesses across Jilin Province.

I have also received the National Scholarship (top 0.4% of 619 students) and the Sunny Optical Technology (Group) Company Scholarship (top 1% of 619 students), among other distinctions.