Wancong (Kevin) Zhang

I am a PhD student at NYU Courant, advised by Yann LeCun. My research focuses on world models, planning, and representation learning, with the goal of building agents that can learn from observation and interact with the world in an open-ended way.

I am currently especially interested in hierarchical world models and planning across multiple temporal scales, where structure in the model can enable better long-horizon reasoning, more efficient control, and stronger generalization from offline or reward-free data. More broadly, I am interested in self-supervised learning for decision making, computer vision, and embodied intelligence.

Email  /  CV  /  Google Scholar  /  Github

News

• Apr 2026 — Released our new paper on arXiv Hierarchical Planning with Latent World Models.
• Sep 2025 — Paper accepted to NeurIPS 2025 - Planning with Latent Dynamics Models.
• Jul 2025 — Our paper on Planning with Latent Dynamics Models received the Best Paper Award at ICML 2025 Workshop on Physically Plausible World Models.
• Jun 2025 — Started Research Scientist Internship at Meta AI (FAIR) under the mentorship of Nicolas Ballas.
• Sep 2023 — Serving as a Teaching Assistant for Deep Learning (DS-GA 1008)
• Sep 2023 — Began PhD at NYU Courant under the supervision of Yann LeCun.

Research (Updated 04/2026)

Introduces hierarchical planning across latent world models at multiple temporal scales, improving long-horizon zero-shot control while reducing planning-time compute on robotics and simulated control tasks.

This work introduces zero-shot planning using the Joint Embedding Prediction Architecture (JEPA) trained from offline trajectories, demonstrating its strengths in generalization, trajectory stitching, and data efficiency compared to traditional offline RL.

Presents an efficient probing benchmark to evaluate the fitness of unsupervised visual representations for reinforcement learning (RL). Applied it to systematically improve pre-existing SSL recipes for RL.

Showcases an industrial-scale end-to-end Automatic Speech Recognition model trained on 570k hours of speech audio data using Noisy Student. It achieves competitive word error rates against larger and more computationally expensive models.

Adapts the computer vision data augmentation technique MixUp to the natural language domain, reducing calibration error of transformers for sentence classification by up to 50%.

Establishes a pre-training benchmark protocol for electronic health record (EHR) data.

Teaching

• Designed and hosted the final course competition (GitHub)