I was invited to give a talk at the AI Seminar at HKUST (Guangzhou) in November 2026.
The talk, World Models for Planning: From Code World Models to Temporal-Distance JEPA, is about using a world model to plan. A planner needs to know what will happen next, and which future is closer to the goal. The deck covers three papers:
- PatchWorld turns a text environment into a Python program, then fixes the program when it disagrees with past trajectories. After that, planning is just running the program. (paper, code)
- VisualPatchWorld does this for visual control. The shape of the dynamics matters: a program can fit the next step and still get the physics wrong. Contact is the hard part. (paper, code)
- Temporal-Distance-JEPA keeps a neural world model and learns the score used in planning from the order of states in the demonstrations. (paper, code)
Slides: Download PDF
Thanks to the organizers for the invitation.