I completed my undergraduate degree in Mathematics at New York University (NYU). I am currently a research intern at the Shanghai Academy of AI for Science (SAIS), working on JEPA world models for efficient and long-horizon planning.
My interests lie in self-supervised visual representation learning and how learned representations transfer to downstream tasks. I am particularly interested in JEPA world models.
Studying and reproducing JEPA world models, from visual representations and latent dynamics to planning. I am investigating computational bottlenecks in MPC, exploring more efficient planning algorithms, and studying hierarchical models for prediction and planning at multiple temporal scales.
Built a 3.63M-parameter, action-conditioned JEPA model on 2.5M offline trajectories in a two-room environment. Used Barlow Twins warmup, MSE alignment, and encoder freezing to prevent collapse, and evaluated generalization to new layouts and longer rollouts.
Ranked second overall and achieved the best model performance in the course competition.
CVPR OpenSUN3D Workshop, 2026
Co-authored a data-centric survey of 3D representations, datasets, and learning paradigms across reconstruction, generation, and embodied intelligence. Analyzed 50 representative datasets and co-maintain the companion repository, with a focus on world models.
report / slides / repository
Presented Leung and Barron’s Information Theory and Mixing Least-Squares Regressions (2006) and derived the key oracle inequality for exponential-weight model averaging. Coursework covered information-theoretic tools for minimax risk, sample complexity, and statistical estimation.
Presented Fearnhead et al.’s Piecewise Deterministic Markov Processes for Continuous-Time Monte Carlo (2018), worked through generator and invariance derivations, and implemented a Zig–Zag sampler with event-time thinning on a 2D Gaussian target.