World models · JEPA · Planning

After 600 experiments, here's what I discovered about JEPA world models

By Haddy Alchaer · November 2026

Solving long-horizon tasks without any extra training

0/24on target · step 0

Vanilla LeWM can be augmented to solve long-horizon physics tasks on out-of-distribution goals. This method uses transformations on a memory graph to generate on-the-fly subgoals without any additional training. Shown here are some of the hardest tasks it solved.

There's a big debate in AI right now about what a world model should be. Most of the attention goes to generative world models, which predict the future by generating it, usually in pixels. Many scientists, including Yann LeCun, argue that this is wasted effort. Most of the pixels don't matter for deciding what to do, and many of them can't be predicted at all. [1]