I was invited to give a guide talk at the CCF YOCSEF headquarters technical forum on September 12, 2026, at the Institute of Computing Technology, Chinese Academy of Sciences. The forum theme was how world models can capture multiple dimensions of the world—from physical environments to mental and social modeling, simulation, and verification.

My talk, 《代码世界模型：把环境动力学写成可执行的程序》 (Code World Models: Writing Environment Dynamics as Executable Programs), argued for representing dynamics as programs that an LLM can propose, filter, and repair, while the programs themselves carry prediction and planning.

The deck mainly covers two lines of work:

1. PatchWorld — inducing inspectable, replayable Python world models from offline agent trajectories via gradient-free patching (paper, code).
2. VisualPatchWorld — learning executable code world models for visual control, combining structure discovery, parameter fitting, and model-predictive planning (paper, code).

Thanks to the organizers and fellow speakers for a sharp discussion on what world models should model, how multi-dimensional state should evolve, and what counts as evidence of “understanding.”

Slides: Download PDF