Haddy Alchaer · 2026-11-11 · notable
600 JEPA world-model experiments — subgoals, not physics, limit planning
In short
Haddy Alchaer ran about 600 experiments with the LeWM JEPA world model on Push-T. Its learned predictor plans as well as true physics, while adding subgoals lifts long-horizon success from 22% to 80%.
Haddy Alchaer · 2026-11-11 · 3 sources checked
A long hands-on study finds that a JEPA world model's planner, not its predictor, is the real bottleneck.
Key specs
| Long horizon success | 22% → 80% with subgoals |
|---|
What is it?
Haddy Alchaer's post reports on about 600 experiments with LeWM, a JEPA world model from Yann LeCun's group that predicts the future as embeddings instead of pixels. All tests use the Push-T robot task and the published model, with no retraining.
How does it work?
On 250 tasks with a goal 75 steps away, the learned predictor succeeded 78% of the time against 77% for the simulator's exact physics. Planning all 75 steps at once gave only 22%; a subgoal every 25 steps raised it to 80%. A memory graph built from about 460,000 embedded demonstration states, with a cost on each jump, reached 85%.
Why does it matter?
The study points researchers at planning and goal choice, not better predictors, as the next problem for JEPA-style world models. It also shows the standard test is easy: LeWM scored 68% on the official 50 episodes but 44% on 250 fresh ones, so reported Push-T numbers may be inflated.
Who is it for?
world-model and robot-learning researchers